Taxonomic diversity-based domain interaction prediction

Identification of protein domain-domain interactions (DDIs) is an essential step in understanding proteins’ functional and structural roles. MirrorTree is a DDI prediction method that is based on the principle of interacting proteins’ co-evolution. However, this method is sensitive to taxonomic diversity and evolutionary span within the two protein homolog sets compared to predict DDI. In this work, we propose a new MirrorTree-based DDI prediction method, namely Taxonomic Diversity-based Domain Interaction Prediction (TAXDIP). TAXDIP improves the MirrorTree method by adding a sampling step that favors representation of higher-level taxonomic ranks (e.g. family over species) in two protein homolog sets prior to their comparison. This additional step ensures increased evolutionary span within protein homolog sets. TAXDIP is first assessed using a set containing 6,514 positive (interacting) domain pairs and a negative (non-interacting) set of equal size containing randomly generated domain pairs with no known interactions. TAXDIP achieved 71.0% sensitivity and 63.0% specificity on this set. Next, a benchmark-set containing 500 interacting and 500 non-interacting domain pairs is used to compare the performance of TAXDIP against DDI prediction methods ME and RDFF. TAXDIP showed better sensitivity and specificity than RDFF. While TAXDIP’s sensitivity is better than ME, its specificity remained below ME. In conclusion, TAXDIP, with its performance, is a viable alternative to existing prediction methods. Furthermore, given TAXDIP’s true predictions are overlapping with, and furthermore, complementing other DDI prediction methods, TAXDIP has a strong position in becoming part of a meta-DDI prediction method that combines multiple methods to build a consensus prediction