BACKGROUND: Empirical scoring functions have proven useful in protein structure modeling. Most such scoring functions depend on protein side chain conformations. However, backbone-only scoring functions do not require computationally intensive structure optimization and so are well suited to protein design, which requires fast score evaluation. Furthermore, scoring functions that account for the distinctive relative position and orientation preferences of residue pairs are expected to be more accurate than those that depend only on the separation distance. RESULTS: Residue pair scoring functions for fixed backbone protein design were derived using only backbone geometry. Unlike previous studies that used spherical harmonics to fit 2D angular distributions, Gaussian Mixture Models were used to fit the full 3D (position only) and 6D (position and orientation) distributions of residue pairs. The performance of the 1D (residue separation only), 3D, and 6D scoring functions were compared by their ability to identify correct threading solutions for a non-redundant benchmark set of protein backbone structures. The threading accuracy was found to steadily increase with increasing dimension, with the 6D scoring function achieving the highest accuracy. Furthermore, the 3D and 6D scoring functions were shown to outperform side chain-dependent empirical potentials from three other studies. Next, two computational methods that take advantage of the speed and pairwise form of these new backbone-only scoring functions were investigated. The first is a procedure that exploits available sequence data by averaging scores over threading solutions for homologs. This was evaluated by applying it to the challenging problem of identifying interacting transmembrane alpha-helices and found to further improve prediction accuracy. The second is a protein design method for determining the optimal sequence for a backbone structure by applying Belief Propagation optimization using the 6D scoring functions. The sensitivity of this method to backbone structure perturbations was compared with that of fixed-backbone all-atom modeling by determining the similarities between optimal sequences for two different backbone structures within the same protein family. The results showed that the design method using 6D scoring functions was more robust to small variations in backbone structure than the all-atom design method. CONCLUSIONS: Backbone-only residue pair scoring functions that account for all six relative degrees of freedom are the most accurate and including the scores of homologs further improves the accuracy in threading applications. The 6D scoring function outperformed several side chain-dependent potentials while avoiding time-consuming and error prone side chain structure prediction. These scoring functions are particularly useful as an initial filter in protein design problems before applying all-atom modeling.

BACKGROUND: The binding of peptide fragments of extracellular peptides to class II MHC is a crucial event in the adaptive immune response. Each MHC allotype generally binds a distinct subset of peptides and the enormous number of possible peptide epitopes prevents their complete experimental characterization. Computational methods can utilize the limited experimental data to predict the binding affinities of peptides to class II MHC. RESULTS: We have developed the Regularized Thermodynamic Average, or RTA, method for predicting the affinities of peptides binding to class II MHC. RTA accounts for all possible peptide binding conformations using a thermodynamic average and includes a parameter constraint for regularization to improve accuracy on novel data. RTA was shown to achieve higher accuracy, as measured by AUC, than SMM-align on the same data for all 17 MHC allotypes examined. RTA also gave the highest accuracy on all but three allotypes when compared with results from 9 different prediction methods applied to the same data. In addition, the method correctly predicted the peptide binding register of 17 out of 18 peptide-MHC complexes. Finally, we found that suboptimal peptide binding registers, which are often ignored in other prediction methods, made significant contributions of at least 50% of the total binding energy for approximately 20% of the peptides. CONCLUSIONS: The RTA method accurately predicts peptide binding affinities to class II MHC and accounts for multiple peptide binding registers while reducing overfitting through regularization. The method has potential applications in vaccine design and in understanding autoimmune disorders. A web server implementing the RTA prediction method is available at http://bordnerlab.org/RTA/.

BACKGROUND: The binding of peptide fragments of antigens to class II MHC is a crucial step in initiating a helper T cell immune response. The identification of such peptide epitopes has potential applications in vaccine design and in better understanding autoimmune diseases and allergies. However, comprehensive experimental determination of peptide-MHC binding affinities is infeasible due to MHC diversity and the large number of possible peptide sequences. Computational methods trained on the limited experimental binding data can address this challenge. We present the MultiRTA method, an extension of our previous single-type RTA prediction method, which allows the prediction of peptide binding affinities for multiple MHC allotypes not used to train the model. Thus predictions can be made for many MHC allotypes for which experimental binding data is unavailable. RESULTS: We fit MultiRTA models for both HLA-DR and HLA-DP using large experimental binding data sets. The performance in predicting binding affinities for novel MHC allotypes, not in the training set, was tested in two different ways. First, we performed leave-one-allele-out cross-validation, in which predictions are made for one allotype using a model fit to binding data for the remaining MHC allotypes. Comparison of the HLA-DR results with those of two other prediction methods applied to the same data sets showed that MultiRTA achieved performance comparable to NetMHCIIpan and better than the earlier TEPITOPE method. We also directly tested model transferability by making leave-one-allele-out predictions for additional experimentally characterized sets of overlapping peptide epitopes binding to multiple MHC allotypes. In addition, we determined the applicability of prediction methods like MultiRTA to other MHC allotypes by examining the degree of MHC variation accounted for in the training set. An examination of predictions for the promiscuous binding CLIP peptide revealed variations in binding affinity among alleles as well as potentially distinct binding registers for HLA-DR and HLA-DP. Finally, we analyzed the optimal MultiRTA parameters to discover the most important peptide residues for promiscuous and allele-specific binding to HLA-DR and HLA-DP allotypes. CONCLUSIONS: The MultiRTA method yields competitive performance but with a significantly simpler and physically interpretable model compared with previous prediction methods. A MultiRTA prediction webserver is available at http://bordnerlab.org/MultiRTA.

BACKGROUND: Many integral membrane proteins, like their non-membrane counterparts, form either transient or permanent multi-subunit complexes in order to carry out their biochemical function. Computational methods that provide structural details of these interactions are needed since, despite their importance, relatively few structures of membrane protein complexes are available. RESULTS: We present a method for predicting which residues are in protein-protein binding sites within the transmembrane regions of membrane proteins. The method uses a Random Forest classifier trained on residue type distributions and evolutionary conservation for individual surface residues, followed by spatial averaging of the residue scores. The prediction accuracy achieved for membrane proteins is comparable to that for non-membrane proteins. Also, like previous results for non-membrane proteins, the accuracy is significantly higher for residues distant from the binding site boundary. Furthermore, a predictor trained on non-membrane proteins was found to yield poor accuracy on membrane proteins, as expected from the different distribution of surface residue types between the two classes of proteins. Thus, although the same procedure can be used to predict binding sites in membrane and non-membrane proteins, separate predictors trained on each class of proteins are required. Finally, the contribution of each residue property to the overall prediction accuracy is analyzed and prediction examples are discussed. CONCLUSION: Given a membrane protein structure and a multiple alignment of related sequences, the presented method gives a prioritized list of which surface residues participate in intramembrane protein-protein interactions. The method has potential applications in guiding the experimental verification of membrane protein interactions, structure-based drug discovery, and also in constraining the search space for computational methods, such as protein docking or threading, that predict membrane protein complex structures.

MOTIVATION: Specific non-covalent binding of metal ions and ligands, such as nucleotides and cofactors, is essential for the function of many proteins. Computational methods are useful for predicting the location of such binding sites when experimental information is lacking. Methods that use structural information, when available, are particularly promising since they can potentially identify non-contiguous binding motifs that cannot be found using only the amino acid sequence. Furthermore, a prediction method that can utilize low-resolution models is advantageous because high-resolution structures are available for only a relatively small fraction of proteins. RESULTS: SitePredict is a machine learning-based method for predicting binding sites in protein structures for specific metal ions or small molecules. The method uses Random Forest classifiers trained on diverse residue-based site properties including spatial clustering of residue types and evolutionary conservation. SitePredict was tested by cross-validation on a set of known binding sites for six different metal ions and five different small molecules in a non-redundant set of protein-ligand complex structures. The prediction performance was good for all ligands considered, as reflected by AUC values of at least 0.8. Furthermore, a more realistic test on unbound structures showed only a slight decrease in the accuracy. The properties that contribute the most to the prediction accuracy of each ligand were also examined. Finally, examples of predicted binding sites in homology models and uncharacterized proteins are discussed. AVAILABILITY: Binding site prediction results for all PDB protein structures and human protein homology models are available at http://sitepredict.org/.

BACKGROUND: Protein-protein interactions are ubiquitous and essential for all cellular processes. High-resolution X-ray crystallographic structures of protein complexes can reveal the details of their function and provide a basis for many computational and experimental approaches. Differentiation between biological and non-biological contacts and reconstruction of the intact complex is a challenging computational problem. A successful solution can provide additional insights into the fundamental principles of biological recognition and reduce errors in many algorithms and databases utilizing interaction information extracted from the Protein Data Bank (PDB). RESULTS: We have developed a method for identifying protein complexes in the PDB X-ray structures by a four step procedure: (1) comprehensively collecting all protein-protein interfaces; (2) clustering similar protein-protein interfaces together; (3) estimating the probability that each cluster is relevant based on a diverse set of properties; and (4) combining these scores for each PDB entry in order to predict the complex structure. The resulting clusters of biologically relevant interfaces provide a reliable catalog of evolutionary conserved protein-protein interactions. These interfaces, as well as the predicted protein complexes, are available from the Protein Interface Server (PInS) website (see Availability and requirements section). CONCLUSION: Our method demonstrates an almost two-fold reduction of the annotation error rate as evaluated on a large benchmark set of complexes validated from the literature. We also estimate relative contributions of each interface property to the accurate discrimination of biologically relevant interfaces and discuss possible directions for further improving the prediction method.

MOTIVATION: Evolutionary conservation estimated from a multiple sequence alignment is a powerful indicator of the functional significance of a residue and helps to predict active sites, ligand binding sites, and protein interaction interfaces. Many algorithms that calculate conservation work well, provided an accurate and balanced alignment is used. However, such a strong dependence on the alignment makes the results highly variable. We attempted to improve the conservation prediction algorithm by making it more robust and less sensitive to (1) local alignment errors, (2) overrepresentation of sequences in some branches and (3) occasional presence of unrelated sequences. RESULTS: A novel method is presented for robust constrained Bayesian estimation of evolutionary rates that avoids overfitting independent rates and satisfies the above requirements. The method is evaluated and compared with an entropy-based conservation measure on a set of 1494 protein interfaces. We demonstrated that approximately 62% of the analyzed protein interfaces are more conserved than the remaining surface at the 5% significance level. A consistent method to incorporate alignment reliability is proposed and demonstrated to reduce arbitrary variation of calculated rates upon inclusion of distantly related or unrelated sequences into the alignment.