Improving positive unlabeled learning algorithms for protein interaction prediction

In binary classification for protein interaction prediction, labeling two proteins as negative (not interacting) is a hard task. This problem is caused by the difficulty of obtaining two training samples that would never interact. Furthermore, the protein interaction databases do not include negative samples, even if the samples have been shown to be non-interacting. The aforementioned difficulty in obtaining true negative samples created a need for learning algorithms that does not use negative samples. This study aims to improve upon two well-performing positive unlabeled learning algorithms, AGPS and Roc-SVM for protein interaction prediction. Two extensions to these algorithms is proposed; the first one is to use Random Forests as the classifier instead of support vector Machines (AGPS-RF and Roc-RF) and the second is to combine the results of AGPS and Roc-SVM using a voting system (Hybrid Algorithm). After these two approaches are implemented, the results were compared to the original algorithms as well as two well-known learning algorithms, ARACNE and CLR. In the tests and comparisons, both Random Forest algorithms and the Hybrid algorithm performed well against the original SVM-classified ones. The improved Roc-RF and Hybrid Algorithms also performed well against ARACNE and CLR.