A top-down approach for finding interaction detection methods

Abstract

Background: The experimental methods used to detect protein interactions are an essential part of the evidence for the interaction. The BioCreative 3 PPI-IMT task consists of determining which interaction detection methods were used to detect protein interactions in a given full text article. Results: We created a machine-learning based system where each interaction detection method was modeled by one classifier. The classifiers were trained and applied at the document level to provide the overall determination of whether a given method was used in the document. However, we also apply the same classifiers at the sentence level to determine which sentence from the document best supports the interaction detection method. The highest macro-averaged f-measure we achieved in our cross-validation experiments was 0.6278, and the highest AUC iP/R we achieved was 0.6217.