Prediction of protein surface accessibility with information theory

A High Performing Tool for Residue Solvent Accessibility Prediction

“…Competitors use very different approaches to predict the exposure state of residues. For each such approach we selected the best performing tool: [6] for the Information Theory (IT) approach, [8] for Probability Profiles (PP), SARpred [14] for Neural Networks (NN), RSA-PRP [20] for Support Vector Regression (SVR), [23] for a combination of Linear Regression and Support Vector Regression (LR+SVR), and SABLE [13] for a combination of Neural Networks and Linear Regression (NN+LR). We did not compare our tool against those that used a Real Values approach [33,21,15] (including the look-up table approach by Carugo et al [5]), as these are not binary classifiers, which makes output comparison not straightforward.…”

“…This dataset consists of 215 non-homologous protein chains (50878 residues) with no more than 25% pairwise-sequence identity and crystallographic resolution < 2.5Å [6].…”

“…Depending on this percent value, amino acids are then classified as either buried or exposed on a binary basis, or with discrete exposure levels in case of multiple threshold systems [4]. Several different approaches have been proposed to cope with the solvent accessibility problem: Information Theory [5,6], Bayesian Statistics [7], Probability Profiles [8], Neural Networks [4,9,11,12,13,14,15], Linear Regression [16,17], Support Vector Machines [18,19], Support Vector Regression [20], Look-up Tables [21], meta-methods [22] and many others [23]. However, exploiting sequence similarity to known structures, namely sequence homology, proved to be a substantial improvement strategy for all these methods, both for secondary structure and Solvent Accessibility prediction [9,24].…”

Identification of interface residues in protease-inhibitor and antigen-antibody complexes: a support vector machine approach

“…Knowledge of surface topology and the geometric neighbors of residues used in the other studies were not used in our study. Several authors have reported success in predicting surface residues from the amino acid sequence [2,10,12,19,20,21]. This raises the possibility of first predicting surface residues based on sequence information, and then using the predicted surface residue information to predict the interaction sites using an SVM classifier.…”

“…After the removal of redundant proteins and proteins with fewer than ten residues, we obtained a data set of 115 proteins belonging to six different categories of complexes. The six categories and the number of proteins in each category are: antibody-antigen (31), protease-inhibitor (19), enzyme complexes (14), large protease complexes (8), G-proteins, cell cycle, signal transduction (22), and miscellaneous (21). In the study described here, we focused on the proteins from two categories: 19 proteins from protease-inhibitor complexes and 31 proteins from antibodyantigen complexes (the protein list is available at http://www.public.iastate.edu/~chhyan/isda2003/sup.htm).…”