Input-Output Kernel Regression applied to protein-protein interaction network inferenc
Carlos Maycas Nadal · 2015
The study of protein-protein interaction networks has received a lot of attention by the research community lately. However, the experimental studies to reconstruct this kind of structures are expensive. Consequently, several machine learning approaches have been developed that automatically infer PPI networks. In this work I present the implementation and analysis of the Input-Output Kernel Regression (IOKR) developed by [9, 10] to compute the inference using various experimental data sets. IOKR is based on the learning of an output kernel that let us apply regression models on a feature space where we can compute the similarity of pairs of proteins to infer the existence of interactions. Furthermore, this approach extends the Kernel Ridge Regression to a semi-supervised approach where the inference turns into a matrix completion. The Multiple Kernel Learning is applied on the input side to deal with the different data sources. Finally, I compare the performance of the implementation with other supervised approaches for the inference of PPI networks.