Reducing protein-protein interaction features via scalable approaches

Abstract EN

Prediction of protein-protein interaction may lead to predict discovering complex relations between proteins, which is the nucleus for system biology.

Moreover, the ability to predict which proteins in the cell interact with other proteins can help significantly in identifying the function of newly discovered proteins.

To overcome the dimensionality problem of protein-protein interaction data, one has to select which interaction features are required.

In the current work, we propose two scalable approaches : FSSA (Feature Selection Scalable Approach), a scalable approach of RELIEF and Feature Dimension Reduction Filter (FDRF) algorithms and MFSSA (Multi-class Feature Selection Scalable Approach), an integrated approach between FDRF with Relief-F.

FSSA and MFSSA encompass the following characteristics : 1) select relevant features that describe protein-protein interactions and 2) handle multi-class problem of classifying protein-protein interaction types as well : physical interactions, complex interactions, and negative interactions.

To test the performance of both proposed feature selection approaches, we use the following features of 10,000 yeast proteins in case of FSSA and 15,000 yeast proteins in case of MFSSA, where data is available at BIND : homologs in other species, mutual clustering coefficients : Jaccard, Meet-min, and Geometry coefficients, Gene Ontology (GO) annotations: Gene Molecular Function, Gene Biological Process, and Gene Cellular Component.