Fig. 2

Summary of TARA-TS and our evaluation framework. a TARA-TS aims to align two networks (in our study, yeast and human PPI networks). Besides the networks, TARA-TS also uses sequence similar yeast-human protein pairs as anchor links. See section “Methods – Data”. b From the networks and anchor links, TARA-TS builds an integrated yeast-human network and extracts integrated topology- and sequence-based features of node (protein) pairs. See section Methods – TARA-TS’s feature extraction methodology”. c Given the features, TARA-TS trains a classifier on a training set to learn what features distinguish between functionally related and functionally unrelated node pairs, and then the classifier is evaluated on a testing set. To perform this classification, yeast-human node pairs are labeled. If the two nodes in a given pair are functionally related (intuitively, share GO terms), they are labeled with the positive class; if they are functionally unrelated, they are labeled with the negative class. See section “Methods – Data”. Then, the set of labeled node pairs is split into training and testing sets to perform the classification. Only if classification accuracy is high, i.e., if TARA-TS accurately predicts functionally (un)related nodes to be functionally (un)related, does it make sense to use TARA-TS to create an alignment for protein functional prediction. d Node pairs from the testing set that are predicted as functionally related are taken as TARA-TS’s alignment. Note that relying on testing data only to create an alignment avoids any circular argument. See section Methods – TARA-TS’s feature extraction methodology”. e Any alignment, of TARA-TS or an existing NA method such as PrimAlign and TARA, can be given to a protein functional prediction framework to predict protein-GO term annotations. Then, the different methods’ alignments are evaluated in terms of their prediction accuracy (we also evaluate their running times). See section “Methods – Using an alignment for protein functional prediction”