Here's a description of this stage quoting from the project report:

Based on the all-features performance assessment, we selected 142 DWPCs to compute on all observations (all 209,168 compound–disease pairs). The feature selection was designed to remove uninformative features (according to permutation) and guard against edge-dropout contamination. Third, we included 14 degree features, which assess the degree of a specific metaedge for either the source compound or target disease.