Robust hierarchical co-clustering for exploring toxicogenomic biomarkers and their chemical regulators

Abstract

Abstract Toxicity measurement of doses of chemicals (DCs) is one of the most important tasks in toxicology studies and the drug discovery and development process. In this issue, toxicogenomic biomarkers are now playing a vital role in measuring the toxicity of DCs. Differentially expressed genes (DEGs) between DCs-treatment and control groups are considered toxicogenomic biomarkers, and associated chemicals are the regulators of DEGs. The co-clustering technique is now used extensively in toxicogenomic research to investigate co-clusters between genomic biomarkers and their chemical regulators. In the literature, there are few approaches to exploring co-clusters. The hierarchical co-clustering (HCoClust) approach is faster, simpler, and more flexible. Nevertheless, it is not robust against outlier data and there is no instruction about separating upregulatory or downregulatory co-clusters, a crucial goal of toxicogenomic data analysis. Therefore, in this article, we proposed a robust HCoClust (rHCoClust) approach and developed an r-package called “rhcoclust” for its implementation. Simulation results showed that the conventional HCoClust and the proposed rHCoClust performed equally well in detecting co-clusters in the absence of outliers, while rHCoClust performed much better than HCoClust in the presence of outliers. However, rHCoClust outperformed the bi-clustering approaches in detecting co-clusters, since bi-clustering methods only work when row and column clusters are equal, and they have no criterion for detecting upregulatory and downregulatory co-clusters. Then rHCoClust was compared with HCoClust through real data analysis and found that rHCoClust performed better than HCoClust. In the case of real data analysis, the proposed method rHCoClust identified top-ranked two DEGs-clusters (GSTA5, MGST2, GCLC, GCLM, G6PD) and (EHHADH, CYP4A1, ANGPT14, CPT1A) that were significantly expressed by the influence of top-ranked two DCs-clusters (acetaminophen_High _24.hr, nitrofurazone_High_24.hr, methapyrilene_High_24.hr) and (WY.14643_High_24.hr, clofibrate_High_24.hr, gemfibrozil_High_24.hr, benzbromarone_High_24.hr, aspirin_High_24.hr) through the glutathione metabolism (GMP) and PPAR signaling pathway (PPAR-SP) respectively. The literature review also supported these results. Thus, the proposed method would be useful to explore toxicogenomic biomarkers and their chemical regulators from the robustness point of view.