Amino-acid residue association models for large scale protein-protein interaction prediction

The computational prediction of protein-protein interactions (PPI) is an essential complement to direct experimental evidence. Traditional approaches rely on less available or computationally predicted surface properties, show database-specific performances and are computationally expensive for large-scale datasets. Several sensitivity and specificity issues remain. Here, we report a novel method based on 'Amino-acid Residue Associations' (ARA) among interacting proteins which utilizes the accurate and easily available primary sequence. Large scale PPI datasets for six model species (from E. coli to human) were studied. The ARA method shows up to 73%sensitivity and 78% specificity. Furthermore, the method performs remarkably well in terms of stability and generalizability. The performance of ARA method benchmarked against existing prediction techniques shows performance improvement upto 25%. Ability of ARA method to predict PPI across species and across databases is also demonstrated. Overall, the ARA method provides a significant improvement over existing ones in correctly identifying large scale protein-protein interactions,irrespective of the data resource, network size or organism.

The MATLAB code for ARA approach will be made available upon request.