Pathway-based gene selection for disease classification

The identification of disease biomarkers from genetic data such as high-throughput transcriptional profiling screens has attracted a great deal of recent interest due to relevance in prognostication and drug discovery. Biomarker discovery can be modelled as a feature selection problem that aims to find the most discriminating features (genes) for accurate disease classification e.g. healthy vs. diseased samples. Typical feature selection algorithms identify individual genes, and the disease discrimination power of each gene is considered separately. In this paper, we propose a gene selection method incorporating prior biological knowledge about gene pathways to find a group(s) of strongly correlated genes to accurately discriminate complex as well as simple diseases. The proposed method involves a ranking process to identify the most relevant biological pathways in a microarray dataset. A specified number of differentially- expressed genes from relevant pathways is then selected for accurate disease classification. The advantage of this method is that it searches for a group of strongly correlated genes rather than individual genes. We argue that selecting a group of informative and correlated genes from disease-associated pathways is particularly relevant to the disease type or grade. To evaluate the performance of our method, we compare it to five well-known feature selection and ranking methods using two classifiers: K-Nearest Neighbour (KNN) and Support Vector Machine (SVM). Using publicly-available microarray datasets, we found that our algorithm outperforms other methods in terms of disease classification accuracy. Moreover, we were able to reduce the number of genes required to accurately discriminate disease states.