Propensity vectors of low‐ASA residue pairs in the distinction of protein interactions

We introduce low‐ASA residue pairs as classification features for distinguishing the different types of protein interactions. A low‐ASA residue pair is defined as two contact residues each from one chain that have a small solvent accessible surface area (ASA). This notion of residue pairs is novel as it first combines residue pairs with the O‐ring theory, an influential proposition stating that the binding hot spots at the interface are often surrounded by a ring of energetically less important residues. As binding hot spots lie in the core of the stability for protein interactions, we believe that low‐ASA residue pairs can sharpen the distinction of protein interactions. The main part of our feature vector is 210‐dimensional, consisting of all possible low‐ASA residue pairs; the value of every feature is determined by a propensity measure. Our classification method is called OringPV, which uses propensity vectors of protein interactions for support vector machine. OringPV is tested on three benchmark datasets for a variety of classification tasks such as the distinction between crystal packing and biological interactions, the distinction between two different types of biological interactions, etc. The evaluation frameworks include within‐dataset, cross‐dataset comparison, and leave‐one‐out cross‐validation. The results show that low‐ASA residue pairs and the propensity vector description of protein interactions are truly strong in the distinction. In particular, many cross‐dataset generalization capability tests have achieved excellent recalls and overall accuracies, much outperforming existing benchmark methods. Proteins 2010. © 2009 Wiley‐Liss, Inc.