A Robust Approach for Identification of Cancer Biomarkers and Candidate Drugs

Background: Identification of cancer biomarkers that are differentially expressed (DE) under two biological conditions is an important task in many microarray studies. There exist several methods in the literature in this regards and most of these methods designed especially for unpaired samples, which does not satisfy the requirements of paired samples where the gene expressions are taken from the same patients before and after treatment. Furthermore, the traditional biomarker identification methods based on either p-values or fold change (FC) values. However, sometimes, p-value based results do not comply with FC based results due to the smaller variance of gene expressions. There are some methods that combine both p-values and FC values to solve this problem. But, these methods also show weak performance for small-sample case in presence of outlying expressions. To overcome this problem, in this paper an attempt is made to develop a hybrid robust SAM-FC approach by combining rank of FC values and rank of p-values based on SAM statistic using minimum β-divergence method, which is designed for paired samples. This method introduces a weight function known as β-weight function. This weight function produces larger weights corresponding to usual/normal expressions and smaller weights for unusual/outlying expressions. The β-weight function plays the significant role on the performance of the proposed method. Results: The proposed method uses β-weight function as a measure of outlier detection by setting β=0.2. We unify both classical and robust estimates using β-weight function such that maximum likelihood estimators (MLEs) are used in absence of outliers and minimum β-divergence estimators are used in presence of outliers to obtain reasonable p-values and FC values in the proposed method. We examined the performance of proposed method in a comparison of some popular methods (t-test, SAM, LIMMA, Wilcoxon, WAD, RP and FCROS) using both simulated and real gene expression profiles for both small-and large-sample cases. From the simulation and a real spike in data analysis results we observed that the proposed method outperforms other methods for small-sample case in presence of outliers and it keeps almost equal performance with other robust methods (Wilcoxon, RP and FCROS) otherwise. From a head-and-neck cancer (HNC) dataset the proposed method identified 2 genes (CYP3A4, NOVA1) that are significantly enriched in linoleic acid metabolism, drug metabolism, steroid hormone biosynthesis and metabolic pathways. The survival analysis through Kaplan-Meier curve revealed that combined effect of these 2 genes has prognostic capability and they might be promising biomarker of HNC. Moreover, we retrieved the 12 candidate drugs based on gene interaction from glad4u and drug bank databases. Conclusion The identified drugs showed statistical significance and critical role of the proteins indicate that these proteins might be therapeutic target in cancer. Thus, elucidating the associations between the drugs identified in the present study require further investigations.