Interrogating domain-domain interactions with parsimony based approaches

Abstract

Background: The identification and characterization of interacting domain pairs is an important step towards understanding protein interactions. In the last few years, several methods to predict domain interactions have been proposed. Understanding the power and the limitations of these methods is key to the development of improved approaches and better understanding of the nature of these interactions.

Results: Building on the previously published Parsimonious Explanation method (PE) to predict domain-domain interactions, we introduced a new Generalized Parsimonious Explanation (GPE) method, which (i) adjusts the granularity of the domain definition to the granularity of the input data set and (ii) permits domain interactions to have different costs. This allowed for preferential selection of the so-called "co-occurring domains" as possible mediators of interactions between proteins. The performance of both variants of the parsimony method are competitive to the performance of the top algorithms for this problem even though parsimony methods use less information than some of the other methods. We also examined possible enrichment of co-occurring domains and homo-domains among domain interactions mediating the interaction of proteins in the network. The corresponding study was performed by surveying domain interactions predicted by the GPE method as well as by using a combinatorial counting approach independent of any prediction method. Our findings indicate that, while there is a considerable propensity towards these special domain pairs among predicted domain interactions, this overrepresentation is significantly lower than in the iPfam dataset.

Conclusion: The Generalized Parsimonious Explanation approach provides a new means to predict and study domain-domain interactions. We showed that, under the assumption that all protein interactions in the network are mediated by domain interactions, there exists a significant deviation of the properties of domain interactions mediating interactions in the network from that of iPfam data.

Figure 1

Benchmark pairs among top-scoring predictions.

Figure 2

ROC Curves of GPE and PE.

Figure 3

Comparison of the Positive Predictive Values for several methods relative to the corresponding random performance. The methods are grouped according to the domain definition. Note that performance of Random varies between the groups. GPE* denotes results obtained by projecting supra-domain from the GPE method is back into Pfam domains where the "children" domains inherit the scores from the supra-domain. A more formal comparison method from different groups and relies on counting how often each of them over/under-performed the corresponding random selection and is described in the text. The performance of GPE and PE was identical while their desistance to the next closed method was statistically significant.