Multitask learning for host–pathogen protein interactions

Kshirsagar, Meghana; Carbonell, Jaime; Klein-Seetharaman, Judith

For each protein pair, we compute features similar to the work in Kshirsagar et al. (2012) . Some features use both proteins in the pair, while some others are based on either the host protein or the pathogen protein. While the features used for S.typhi were obtained directly from the authors, those for the other three datasets were derived from the following attributes of proteins available in public databases: protein sequences from Uniprot (UniProt Consortium, 2011), gene ontology from GO database (Ashburner et al., 2000) , gene expression from GEO (Barrett et al., 2011) , properties of human proteins in the human PPI network. Owing to the lack of space, we briefly mention only some of the prominent features here, and encourage the readers to refer to the supplementary for details. The sequence features count the frequency of amino acid-based n-grams or n-mers (for n ¼ 2, 3, 4, 5) in the protein sequence. The GO features count the co-occurrence of host-pathogen GO term combinations. The human PPI network-based features compute various graph properties like node-degree, betweenness-centrality, clustering coefficient of the human protein.