1. Protein-protein interaction (PPI) data is sourced from BioGRID.

2. A network of physically interacting proteins is constructed using the "igraph" R package.

3. Self-loops and duplicate interactions are removed to consider only unique interactions.

4. 20 node and pairwise properties are calculated using the constructed PPI network.

Gene interactions (GIs) from different sources are mapped onto the PPI network.

For a given SL pair X and Y, the network property for the pair (NP(XY)) is calculated as the average of NP(X) and NP(Y).

Training dataset comprises 70% of GIs from CGIDB, BIOGRID, SLDB, and SVDR, along with NOT data.

The remaining 30% serves as the test data.

Model achieves an accuracy of ~87%.

"varImp" function in the "randomForest" package determines discriminatory features.

"RFE" feature selection algorithm identifies essential features.

Topological properties such as "shortest path," "average betweenness," and "average triangle" significantly contribute to model performance.

STRING database provides a broader PPI set with 19624 proteins and 13,715,405 interactions.

Network is built for STRING proteins, and GIs from CGIdb, BioGRID, and SLDB are mapped.

NOT data is constructed similarly as discussed earlier.

Pairwise combinations of proteins are generated, and GIs from all three datasets, along with yeast homologs, are removed.

A large NOT dataset is created, with 10,000 randomly selected pairs.

Network properties of these GIs are calculated and used to train the model.

Model achieves an accuracy of 82.22% for 10-fold cross-validation.

Among 20 network properties, "average betweenness," "average triangle," "average coreness," and "average neighbor2" are the most discriminatory in classifying the three sets of GIs.