A novel method incorporating gene ontology information for unsupervised clustering and feature selection

Abstract

Background: Among the primary goals of microarray analysis is the identification of genes that could distinguish between different phenotypes (feature selection). Previous studies indicate that incorporating prior information of the genes' function could help identify physiologically relevant features. However, current methods that incorporate prior functional information do not provide a relative estimate of the effect of different genes on the biological processes of interest.

Results: Here, we present a method that integrates gene ontology (GO) information and expression data using Bayesian regression mixture models to perform unsupervised clustering of the samples and identify physiologically relevant discriminating features. As a model application, the method was applied to identify the genes that play a role in the cytotoxic responses of human hepatoblastoma cell line (HepG2) to saturated fatty acid (SFA) and tumor necrosis factor (TNF)-alpha, as compared to the non-toxic response to the unsaturated FFAs (UFA) and TNF-alpha. Incorporation of prior knowledge led to a better discrimination of the toxic phenotypes from the others. The model identified roles of lysosomal ATPases and adenylate cyclase (AC9) in the toxicity of palmitate. To validate the role of AC in palmitate-treated cells, we measured the intracellular levels of cyclic AMP (cAMP). The cAMP levels were found to be significantly reduced by palmitate treatment and not by the other FFAs, in accordance with the model selection of AC9.

Conclusions: A framework is presented that incorporates prior ontology information, which helped to (a) perform unsupervised clustering of the phenotypes, and (b) identify the genes relevant to each cluster of phenotypes. We demonstrate the proposed framework by applying it to identify physiologically-relevant feature genes that conferred differential toxicity to saturated vs. unsaturated FFAs. The framework can be applied to other problems to efficiently integrate ontology information and expression data in order to identify feature genes.

Figure 1. The cytotoxicity and ketone body production in response to various treatments.

Confluent HepG2 cells were treated for 24 h with 0.7 mM of the indicated FFA complexed to 4% (w/v) BSA, in the presence or absence of TNF-α (0, 20 or 100 ng/ml). (A) Cytotoxicity of the treatments. The cytotoxicity was measured as the % LDH released, as defined in the methods section. (B) Ketone body production. Acetoacetate and beta-hydroxybutyrate release into the media were measured by enzymatic assays. Ketone body release was calculated as the sum of acetoacetate and beta-hydroxybutyrate release. Data presented as mean±s.d. of three independent experiments. ★, significant FFA effect, p<0.01, #, significant TNF-α effect, p<0.01.

Figure 2. Discrimination of cytotoxic conditions by the IMGO analysis.

The ability of the two-population model for cytotoxicity to distinguish the cytotoxic (high LDH release) conditions was tested.

Figure 3. Effect of FFA-treatments on intracellular cAMP levels.

Cells were treated for 24 h with 0.7 mM of different types of FFA and the levels of intracellular cAMP were measured. H = Control medium, O = 0.7 mM oleate, L = 0.7 mM linoleate, and P = 0.7 mM palmitate. ★, p<0.01 by a two-tailed t-test.