Prediction of human protein-protein interaction by a mixed Bayesian model and its application to exploring underlying cancer-related pathway crosstalk

Abstract

Protein-protein interaction (PPI) prediction method has provided an opportunity for elucidating potential biological processes and disease mechanisms. We integrated eight features involving proteomic, genomic, phenotype and functional annotation datasets by a mixed model consisting of full connected Bayesian (FCB) model and naive Bayesian model to predict human PPIs, resulting in 40 447 PPIs which contain 2740 common PPIs with the human protein reference database (HPRD) by a likelihood ratio cutoff of 512. Then we applied them to exploring underlying pathway crosstalk where pathways were derived from the pathway interaction database. Two pathway crosstalk networks (PCNs) were constructed based on PPI sets. The PPI sets were derived from two different sources. One source was strictly the HPRD database while the other source was a combination of HPRD and PPIs predicted by our mixed Bayesian method. We demonstrated that PCNs based on the mixed PPI set showed much more underlying pathway interactions than the HPRD PPI set. Furthermore, we mapped cancer-causing mutated somatic genes to PPIs between significant pathway crosstalk pairs. We extracted highly connected clusters from over-represented subnetworks of PCNs, which were enriched for mutated gene interactions that acted as crosstalk links. Most of the pathways in top ranking clusters were shown to play important roles in cancer. The clusters themselves showed coherent function categories pertaining to cancer development.

Figure 1.

Prediction performance. (a) ROC curve of different methods for PPI prediction. ‘Orth + Phen + Gene’ represents the combination of orthologue mapping of physical protein interactions from model organisms (Orth), phenotype similarity (Phen) and Genetics interaction (Gene); ‘TF + Coexp’ represents the combination of regulation of common transcriptional factors (TF) and meta-analysis of coexpression (Coexp); ‘Func + Dom + PTM’ represents the combination of shared biological functional annotation (Func), Interaction of domains (Dom), Co-occurrence of post translational modification pair (PTM); ‘All’ represents the integration of all methods. (b) log2 (LR) cutoffs versus precision of PPI prediction.

Figure 2.

Overlap of PPIs in different databases and predicted PPIs. (a) Overlaps of predicted PPIs derived from different LR cutoffs and human PPI databases (HPRD, black bar; BioGRID, dark grey bar; MINT, white bar and IntAct, light grey bar). (b) Overlaps of Predicted PPIs derived from different LR cutoffs and Predicted PPIs with different posterior odds cutoffs in PIPs database; black bar, PIPLR100, dark grey bar, PIPLR400 and light grey bar, PIPLR1000 represent the PPIs datasets derived from the posterior odds cutoffs of 0.25, 1 and 2.5 separately. (c) The number of predicted PPIs under the condition of increasing LR cutoffs.

Figure 3.

Top three clusters of over-represented subnetwork of mixPCN enriched for interactions between mutated genes in brain cancer. The light grey nodes represent pathways which ranked in top 50 crosstalk in over-represented subnetworks by either number of mutated gene pairs or p-value ranking methods; the triangle nodes indicate pathways which did not rank in top 50 crosstalk in over-represented subnetworks by two ranking methods; squares represent pathways which have been demonstrated to participate in brain cancers through interacting with other pathways in the same cluster in more than one publication; diamonds suggest that more than one interaction with this pathway in this cluster can be verified, but no direct evidence for their roles in brain cancer.