Multitask learning for host–pathogen protein interactions

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

(i) The human proteins involved in each interaction dataset are used to obtain the human pathways that are enriched. We use Fisher's test (based on the hypergeometric distribution) to compute the P-value of each pathway. We plot these P-values for each pathway, and for each dataset in the form of a heatmap shown in Figure 4 . The heatmap shows how there are several commonly enriched pathways across the datasets (the black vertical lines spanning all four rows). It also shows the difference in the enrichment for the S.typhi dataset, which comes from small-scale PPI experiments. (ii) We analyze the similarity between the PPIs from various datasets. A natural way to determine similarity is to check if proteins known to interact in one dataset have homologous proteins that are also interacting in another dataset. Such pairs of proteins, also called 'interologs', are defined as a quadruple of proteins A, B, A 0 , B 0 , where A $ B (interaction) and A 0 $ B 0 . Further, A, A 0 are homologs and B, B 0 are also homologs. The number of such interologs existing between the four datasets is shown in Table 2 . To compute homologs of a protein, we used BLASTP sequence alignment with an e-value cutoff of 0.1. As evident from Table 2 , there are few interologs across the bacterial PPIs. None of the high-throughput datasets have an interolog in the small-scale S.typhi dataset. This seems to indicate that interolog-based approaches to compute task similarity are not relevant here. The phenomenon governing the similarity of these host-pathogen interactions is probably at a much higher level, rather than at the level of individual proteins. We explore one such possibility-the 'commonality hypothesis'.