Efficient Two-Stage Model for PredictingProtein-Protein Interactions

Genome sequencing have identified many predictedproteins with unknown biological functions. Protein interactionnetworks link proteins/genes together, provide a global context forinteracting proteins, and enable studies at the systems level.High throughput methods were developed to detect protein-proteininteractions. However, current techniques show high levels of falsepositives and false negatives. They are also labor intensive andexpensive to perform. Computational methods are needed to reducetesting space and enhance testing efficiency. In this work, atwo-stage statistical scoring model was built to assignconfidence scores to fruit fly yeast two-hybrid interactions. Highconfidence predictions are significantly enriched with biologicallymeaningful interactions. Cross-validation showed good prediction performance,which were also validated by independent data sources. Weobtained 24798 new interactions involving 4702 proteins. The work shouldbe especially interesting to researchers in interactome or systemsbiology fields. It also appeals to professionals practicing machinelearning or data mining in computational biology and bioinformatics.