Enhancing the Coevolutionary Signal

Abstract

Popular coevolutionary methods for predicting residue-residue contacts in 3D proteins structure from aligned sequences use an arbitrary cutoff to separate the signal from the noise. These methods, like GREMLIN and PSICOV, rely on a fixed cutoff value from a rank-sorted list of potential contacts. We show that by considering the local signal within the GREMLIN or PSICOV score matrix, we can increase both the accuracy and the number of correctly predicted contacts. Our approach uses a random forest classification scheme and reveals that even a simple Gaussian kernel can improve the contact prediction.