Outline of the proposed method. Firstly, we manually curated a compendium of curated chemosensitivity related genes (CCRGs) from published papers. Then we selected genes on the microarray that had same enriched GO categories and network characteristics with the CCRGs. These genes were considered as candidate CRGs. To get CRGs for each drug, we further filtered the initial drug-candidate CRG network based on PCC of drug-CCRGs. Filter A is based on Gene Ontology. We characterized CCRG using GO enrichment analysis with Fisher Exact Test. We considered three aspects of GO: biological process (BP), molecular function (MF), and cellular component (CC). p represents the enrichment significance. If enriched p value is smaller than 0.01, CCRGs are significantly enriched in the GO term. Moreover, we investigated that whether CCRGs exhibited functional consistency. We compared the functional similarity of CCRG enriched GO terms to randomly selected gene enriched GO terms. We found that CCRG enriched GO terms exhibited higher similarity scores compared to randomly selected genes. Thus, we regarded all genes in the enriched GO terms as candidate CRGs. Filter B is based on protein interaction networks. We analyzed several network features such as degree and betweenness centrality in six PPINs. Degree and betweenness centrality were selected as network features to prioritize CRGs. The green curve represents betweenness centrality of random genes, and the vertical green line is the betweenness centrality of CCRGs. The blue curve represents degree of random genes, and the vertical blue line is the degree of CCRGs. Filter C is based on gene expression. The majority of drug-CCRGs exhibit a low correlation between gene expression and drug activity. We ranked the absolute PCC of all drug-CCRG pairs in ascending order and set the PCC threshold as 5th percentile of all PCCs.
Chen et al. BMC Medical Genomics 2012 5:43 doi:10.1186/1755-8794-5-43