• LRRK2
• machine learning
• coevolution
• protein domains
• interactome

Protein-protein interactions are often mediated by protein domains. Experimentally determining domain-domain interactions (DDIs) is expensive and time-consuming, so developing bioinformatic tools to predict DDIs would be crucial to break down the interaction network to the domain level resolution. In this work, we present a new DDI prediction model which combines various features, including coevolution, mutual information and common GO terms, in a machine learning framework. Predictions were further improved using the information of available 3D templates of the domain pair in publicly available structures. In order to add this feature to the model, we generated a 3D interacting domains catalog based on the InterPro domain definition. Our catalog extends the information provided by public resources such as 3did and could be readily used for structural annotation of interactomic studies.

We then show the application of our DDI predictor to the analysis of an experimentally determined proximity interactome dataset of Parkinson’s disease-related protein LRRK2. The interaction probabilities provided by our model were used to distinguish between interactions associated with LRRK2 enzymatic activity (involving the catalytic core) and interactions involving more evolutionarily recent domains (ARM, ANK and WD40), which are likely involved in the functional specialization of LRRK2.