Protein-Protein Interactions (PPIs) have a key role in almost all biological processes. Experimental methods such as Yeast two-hybrid (Y2H) can detect a set of PPIs. However, the result is often incomplete and characterized by high false positive (negative) rates. To address this problem, many computational methods have been developed to identify PPIs from high-throughput genomic data.
In this thesis, we have evaluated the performance of four computational methods for inferring interacting and non interacting protein pairs from gene expression data. These methods target the learning of Gaussian Graphical Models (GGMs) from high dimensional data such as gene expression data. The methods rely on inverse covariance estimations for fitting GGMs and they have mainly been applied to learn Gene-Gene interactions (GGIs) from expression data.
The methods are Graphical-Lasso, Or-graph, GeneNet, and qpGraph. Graphical-Lasso and Or-graph are based on sparse inverse covariance matrix estimation. GeneNet relies on regularized inverse covariance matrix estimation and qp-graph depends on limited order partial correlation. We apply these methods on Saccharomyces cerevisiae (Yeast) gene expression data for inferring interacting and non-interacting protein pairs. Then we evaluate the performance of the methods using test data that is constructed using STRING and NEGATOME protein databases. Our analysis reveals that qpGraph performs significantly better than the other methods in inferring both interacting and non-interacting protein pairs.
