Combining gene essentiality with feature selection method to explore multi-cancer biomarkers

Biomarker discovery plays an important role in cancer diagnosis and prognosis assessments. The biomarkers that could be applied among different cancer types are highly useful. Although many traditional feature selection algorithms have shown their power on picking discriminative genes, they are incapable of identifying biologically meaningful biomarkers. Here, on the hypothesis that gene essentiality would be disrupted in cancers, we estimated the gene sets with significant essentiality alteration in six cancer types. We found that different cancer types would share some common gene essentiality alterations. Then, the variable combination population analysis (VCPA) algorithm was applied to identify the potential biomarkers from these common genes, which were used to construct prediction models and exhibited satisfactory classification ability (averaged accuracy: 0.9752) among six cancer types. Interestingly, these biomarkers would tend to cluster as a subnetwork and be characterized by high centrality values in the protein-protein interaction network. They were significantly enriched in the cell cycle and DNA replication pathway which are hallmark signatures of cancers. Several biomarkers have been even verified by the literature searching, reported having roles in chromosome instability and aberrantly expressed between cancer/normal samples. An additional comparison analysis between the VCPA and other six feature selection methods in WEKA suggested biomarkers by VCPA perform superior over those by other methods. These results suggested that our method is promising in identifying the potential multi-cancer biomarkers.