Genome-wide Protein-chemical Interaction Prediction

The analysis of protein-chemical reactions on a large scale is critical to understanding the complex interrelated mechanisms that govern biological life at the cellular level. Chemical proteomics is a new research area aimed at genome-wide screening of such chemical-protein interactions. Traditional approaches to such screening involve in vivo or in vitro experimentation, which while becoming faster with the application of high-throughput screening technologies, remains costly and time-consuming compared to in silico methods. Early in silico methods are dependant on knowing 3D protein structures (docking) or knowing binding information for many chemicals (ligand-based approaches). Typical machine learning approaches follow a global classification approach where a single predictive model is trained for an entire data set, but such an approach is unlikely to generalize well to the protein-chemical interaction space considering its diversity and heterogeneous distribution. In response to the global approach, work on local models has recently emerged to improve generalization across the interaction space by training a series of independant models localized to each predict a single interaction. This work examines current approaches to genome-wide protein-chemical interaction prediction and explores new computational methods based on modifications to the boosting framework for ensemble learning. The methods are described and compared to several competing classification methods. Genome-wide chemicalprotein interaction data sets are acquired from publicly available resources, and a series of experimental studies are performed in order to compare the the performance of each method under a variety of conditions.