论文部分内容阅读
Background: Severalstudies, such as the Connectivity Map (CMAP), showed that gene expression profiles could reflect drug perturbationeffects, and can be used in drug screening and drug repositioning.Previously we introduced a concept CoMi (Context-Specific miRNA Activity), and provide a novel perspective on drug mechanisms of action.In this work, we intend to construct a virtual drug screening system and test whether CoMi could be used to capture the common features of approved cancer drugs and discriminate them from other candidate chemicals.Methods & Results: (1)We used correlation analysisby integrating the drugsensitive data (GI50) with gene expression datafrom diverse cancer cell lines in NCI60 datasets, and generated the drug sensitivity gene sets.By calculating the overlap rate between the drug sensitivity gene sets with disregulated gene set in breast cancer, we defined a drug-disease association index andgenerated a pool of 445 breast cancer associated compounds, including 61 positive drugs (approved breast cancer drugs) and 384 negative drugs.(2)Using the most significant differential CoMi features in positive vs.negative drugs, a Na(i)veBayesian classifier could accurately predict the successful breast cancer drugs (AUC 0.86), and its performance was significantly exceed the mRNA signature based drug screening system (AUC 0.75), far superior to naive CMAP screening system (AUC0.56).(3)Further analysis of the network consisted of CoMi with good classification performance,highlighted some high connected miRNA nodes,including important cancer-related miRNAs (mir-495) and tumor suppressormiRNA (mir-520d) ; In addition, a number of high connectivity functional gene setsare also closely related with the biological processes of cancer (for example, cell differentiation).(4)Integrating differential expressed CoMi induced by Paclitaxel with breast cancer specific CoMi, weexaminedits application on classifying drug responsiveness in a breast cancer cohort using paclitaxel as neoadjuvant therapy.We demonstrated that the above CoMi feature could accurately predict the therapeutic outcome (AUC 0.75, pathological complete response (pCR)vs.residual disease (RD)).Conclusion: microRNAs regulation networks and CoMi feature outperformed othergenomic features in a virtual drug screening system, suggests that microRNAs regulation network feature could facilitatein silico cancer drug screening and drug responsiveness signature identification .