Gene network inference using biological homogeneity index based-clustering and constraint-based searching

Abstract

Gene network inference involves exploration of an exponential search space. Initially, network inference utilizes microarray data as a single data source. However, due to microarray data limitations, other biological data is combined with microarray data for network inference. Previous research has produced biological homogeneity measures based on functional annotations from Gene Ontology for various clustering algorithms. Biological homogeneity measures the ability of a clustering algorithm to produce biologically meaningful clusters. Biological Homogeneity Index (BHI) is measured for a range of fc values for fc-means clustering algorithm to find clusters which score the highest homogeneity index. Results are compared using whole dataset, fc-means clusters and fc-means clusters with BHI (fc-means /BHI) approaches. Experimental results have shown that the fc-means clusters produced statistically significant valid number of gene interactions compared to the whole dataset. In comparing the fc-means clusters and fc-means /BHI clusters, the fc-means /BHI clusters produces more valid number of gene interactions for all experiments. Statistical significance test results show that these improvements are too small to be statistically significant. Hence, biological enrichment scores are also used for evaluation. Enrichment scores for fc-means /BHI clusters are better than scores for fc-means clusters. This research employs the constraint-based search algorithm called Grow-Shrink algorithm (GS) in learning the best network structure. Experiments are performed to compare the performance for constraint-based search against scorebased approaches such as the Greedy Search (GRS) and Simulated Annealing (SA). Experimental results prove that GS performs better than GRS and SA in terms of valid interactions number. However, the improvements are too small to be statistically significant. The thesis concludes that using prior biological knowledge can help form biologically meaningful clusters. Using constraint-based search algorithm is also useful for improving the quality of gene network inference.