VDM Verlag Efficient Two-Stage Model for Predicting Protein-Protein Interactions: Machine Learning

Genome sequencing have identified many predicted proteins with unknown biological functions. Protein interaction networks link proteins/genes together, provide a global context for interacting proteins, and enable studies at the systems level. High throughput methods were developed to detect protein-protein interactions. However, current techniques show high levels of false positives and false negatives. They are also labor intensive and expensive to perform. Computational methods are needed to reduce testing space and enhance testing efficiency. In this work, a two-stage statistical scoring model was built to assign confidence scores to fruit fly yeast two-hybrid interactions. High confidence predictions are significantly enriched with biologically meaningful interactions. Cross-validation showed good prediction performance, which were also validated by independent data sources. We obtained 24798 new interactions involving 4702 proteins. The work should be especially interesting to researchers in interactome or systems biology fields. It also appeals to professionals practicing machine learning or data mining in computational biology and bioinformatics.