Generating candidate subnetworks

The primary goal of SPSNet is to identify biological factors that distinguish subpopulations within a sample. Therefore, pathways were chosen to generate subnetworks as they represent the biological processes in an organism, and differences in their functioning contribute to differences within phenotypes. SPSNet does not preclude generating subnetworks from high-quality PPI networks. Both PPI networks and biological pathways can be supplied, even simultaneously, as input to SPSNet (and also to PFSNet). However, in the present manuscript, we do not investigate PPI networks since there are confounding issues when using PPI networks. For example, a PPI network is strictly speaking an artificial assembly of pairwise PPIs: While each individual PPI is a real biological interaction, the subnetwork itself is misleading because e.g. not all partners of a protein in the subnetwork actually simultaneously bind the protein. To ensure a straightforward interpretation and evaluation of our method, we prefer to exclude PPI networks in this manuscript.

The standard PFSNet methodology uses highly expressed genes from each phenotype to induce subnetworks on known biological pathways. However, this technique for generating candidate subnetworks is not suitable for heterogeneous data, as the presence of multiple subpopulations in a phenotype is likely to dilute high expression in any specific subtype. Therefore, we generate subnetworks as in NEA [21]; i.e. we form a subnetwork from each gene and its immediate neighbors in a biological pathway. We filter out subnetworks with less than 5 genes. We generate a total of 5654 such subnetworks from 300 human pathways in PathwayAPI [18].