Background: Colorectal cancer (CRC) could be resected and cured if detected at an early stage. A serum metabolomics-based screening test would represent a rapid and convenient means to identify adenoma and early CRC. However, changes in the metabolome with such minimal disease may be too subtle if the usual high throughput discovery approach were taken. We postulated that adenoma-associated metabolic changes could be better identified by targeting our analysis to metabolites that changed in a progressive manner with sequential stages of disease. Methods: Sera from patients with locoregional (LR CRC) and metastatic (mCRC) disease (stage I n=21, stage II n=30 , stage III n= 35, stage IVa n=67) and matched controls (n=121) were analyzed by gas chromatography-mass spectrometry (GC-MS). After normalization, batch correction and autoscaling, 249 features were selected. Metabolites that characterized CRC were filtered by identifying those metabolites that increased or decreased in a progressive manner with higher disease stage. Subsequently, sera from 31 adenoma patients and 31 matched controls, collected under standardized conditions prior to colonoscopy and polypectomy, were analyzed by GC-MS. "The variable importance” of metabolites identified in the first analysis were studied. Results: Products of fatty acid metabolism (dodecanoic acid, 2-amino butanoic acid, isocaproic acid, hexadecanoic acid) and urea were among metabolites that were altered more markedly in advanced disease stages. A number of metabolites that changed in this progressive fashion were found to change to a more subtle degree in adenoma. Conclusions: A targeted analysis of metabolites that change in a progressive fashion as CRC progresses represents a means to identify the subtle changes in the serum metabolome that accompany adenoma. This demonstrates the feasibility of developing a serum metabolomics-based screening test for CRC. In addition, extension of this selection approach can be used in other applications where minimal disease is challenging to detect with conventional statistical methods.