For each combination of species and the sampling method for negative examples, 
the 40 datasets are numbered 11, 12, 13, 14, 21, 22, 23, 24, 31, ..., 104.

To obtain the predictive performances for the C1, C2 and C3 classes, a protein-protein interaction (PPI) prediction method
is trained with, e.g., yeast_random_train_52.txt, and the trained method is tested on yeast_random_test_C1_52.txt, yeast_random_test_C2_52.txt, and yeast_random_test_C3_52.txt.

To obtain conventional cross-validated predictive performance, a PPI prediction method 
is trained with, e.g., yeast_random_train_CV_52.txt, and the trained method is tested on yeast_random_test_CV_52.txt.

The raw results for the seven methods that were tested in the paper are also posted. You can compute sensitivity, specificity, recall and precision by comparing the raw results with known labels for each protein pair. Please note that, by the program convention, M1 through M3 use -1 for positive protein pairs and 1 for negative protein pairs. M4 through M7, as usual, use 1 for positive protein pairs and -1 for negative protein pairs.

For help, please contact Yungki Park at yungki at mail dot utexas dot edu