Bottom-Up Approach to the Discovery of Clinically Relevant Biomarker Genes: The Case of Colorectal Cancer

Simple Summary This manuscript describes a novel 'bottom-up' approach to the discovery of putative biomarkers of cancer. Unlike traditional 'omics' and 'big data' research approaches, which gather data rather indiscriminately and at a great cost, our approach relies on the knowledge of cell molecular biology. A very small set of known markers are used to 'train' the prediction algorithm, which utilizes accessible and affordable means to analyze and extract relevant information, leading directly to the markers at a fraction of the time and cost of the 'omics' research. To illustrate the capabilities of the proposed approach, we applied the method to colorectal cancer, one of the most common and thoroughly studied cancers. Our method yielded an extended and validated set of 138 putative molecular biomarkers. We further tested 42 of these genes and showed that 41 mRNAs are differentially expressed as we predicted. Our method offers a widely applicable strategy for cancer marker discovery. Traditional approaches to genome-wide marker discovery often follow a common top-down strategy, where a large scale 'omics' investigation is followed by the analysis of functional pathways involved, to narrow down the list of identified putative biomarkers, and to deconvolute gene expression networks, or to obtain an insight into genetic alterations observed in cancer. We set out to investigate whether a reverse approach would allow full or partial reconstruction of the transcriptional programs and biological pathways specific to a given cancer and whether the full or substantially expanded list of putative markers could thus be identified by starting with the partial knowledge of a few disease-specific markers. To this end, we used 10 well-documented differentially expressed markers of colorectal cancer (CRC), analyzed their transcription factor networks and biological pathways, and predicted the existence of 193 new putative markers. Incredibly, the use of a validation marker set of 10 other completely different known CRC markers and the same procedure resulted in a very similar set of 143 predicted markers. Of these, 138 were identical to those found using the training set, confirming our main hypothesis that a much-expanded set of disease markers can be predicted by starting with just a small subset of validated markers. Further to this, we validated the expression of 42 out of 138 top-ranked predicted markers experimentally using qPCR in surgically removed CRC tissues. We showed that 41 out of 42 mRNAs tested have significantly altered levels of mRNA expression in surgically excised CRC tissues. Of the markers tested, 36 have been reported to be associated with aspects of CRC in the past, whilst only limited published evidence exists for another three genes (BCL2, PDGFRB and TSC2), and no published evidence directly linking genes to CRC was found for CCNA1, SHC1 and TGFB3. Whilst we used CRC to test and validate our marker discovery strategy, the reported procedures apply more generally to cancer marker discovery.