Filtering high-throughput protein-protein interaction data using a combination of genomic features

Abstract

Background: Protein-protein interaction data used in the creation or prediction of molecular networks is usually obtained from large scale or high-throughput experiments. This experimental data is liable to contain a large number of spurious interactions. Hence, there is a need to validate the interactions and filter out the incorrect data before using them in prediction studies.

Results: In this study, we use a combination of 3 genomic features -- structurally known interacting Pfam domains, Gene Ontology annotations and sequence homology -- as a means to assign reliability to the protein-protein interactions in Saccharomyces cerevisiae determined by high-throughput experiments. Using Bayesian network approaches, we show that protein-protein interactions from high-throughput data supported by one or more genomic features have a higher likelihood ratio and hence are more likely to be real interactions. Our method has a high sensitivity (90%) and good specificity (63%). We show that 56% of the interactions from high-throughput experiments in Saccharomyces cerevisiae have high reliability. We use the method to estimate the number of true interactions in the high-throughput protein-protein interaction data sets in Caenorhabditis elegans, Drosophila melanogaster and Homo sapiens to be 27%, 18% and 68% respectively. Our results are available for searching and downloading at http://helix.protein.osaka-u.ac.jp/htp/.

Conclusion: A combination of genomic features that include sequence, structure and annotation information is a good predictor of true interactions in large and noisy high-throughput data sets. The method has a very high sensitivity and good specificity and can be used to assign a likelihood ratio, corresponding to the reliability, to each interaction.

Figure 1

Likelihood ratios for genomic features.

Figure 2

ROC curve for the combination of genomic features using 10-fold cross validations. The dotted line shows the empirical ROC curve, while the solid line shows the fitted ROC curve (obtained using JROCFIT). Each point on the ROC curve corresponds to sensitivity and specificity for one or a combination of more than one genomic features. d: interacting Pfam domains; g: similar GO annotations; h: homologous interactions; none: no genomic features. More than one genomic features are indicated by listing the features separated by a '+' sign.

Figure 3

Percentage of interactions predicted true across different high-throughput data sets.

Figure 4

Percentage of interactions predicted true in high and low confidence interactions across different high-throughput data sets.

Figure 5

Some low confidence interactions predicted to be true by our method and confirmed by other publications. The Likelihood ratio for each interaction is indicated. Interactions with a Likelihood ratio greater than 100 are shown with a solid line, while those with a Likelihood ratio less than 10 are shown with a dashed line. (A) Interactions between proteins co-regulating the alternative splicing of Dscam exon 4 in D. menalogaster. (B) Interactions between proteins in the Lsm1-7 complex in S. cerevisiae confirmed by similar interactions found in H. sapiens.