MOPED: method for optimizing physical energy parameters using decoys

Sampling and scoring: a marriage made in heaven

Most structure prediction algorithms consist of initial sampling of the conformational space, followed by rescoring and possibly refinement of a number of selected structures. Here we focus on protein docking, and show that while decoupling sampling and scoring facilitates method development, integration of the two steps can lead to substantial improvements in docking results. Since decoupling is usually achieved by generating a decoy set containing both non-native and near-native docked structures, which can be then used for scoring function construction, we first review the roles and potential pitfalls of decoys in protein-protein docking, and show that some type of decoys are better than others for method development. We then describe three case studies showing that complete decoupling of scoring from sampling is not the best choice for solving realistic docking problems. Although some of the examples are based on our own experience, the results of the CAPRI docking and scoring experiments also show that performing both sampling and scoring generally yields better results than scoring the structures generated by all predictors. Next we investigate how the selection of training and decoy sets affects the performance of the scoring functions obtained. Finally, we discuss pathways to better alignment of the two steps, and show some algorithms that achieve a certain level of integration. Although we focus on protein-protein docking, our observations most likely also apply to other conformational search problems, including protein structure prediction and the docking of small molecules to proteins.

Prediction of hot spot residues at protein-protein interfaces by combining machine learning and energy-based methods

BACKGROUND

Alanine scanning mutagenesis is a powerful experimental methodology for investigating the structural and energetic characteristics of protein complexes. Individual amino-acids are systematically mutated to alanine and changes in free energy of binding (DeltaDeltaG) measured. Several experiments have shown that protein-protein interactions are critically dependent on just a few residues ("hot spots") at the interface. Hot spots make a dominant contribution to the free energy of binding and if mutated they can disrupt the interaction. As mutagenesis studies require significant experimental efforts, there is a need for accurate and reliable computational methods. Such methods would also add to our understanding of the determinants of affinity and specificity in protein-protein recognition.

RESULTS

We present a novel computational strategy to identify hot spot residues, given the structure of a complex. We consider the basic energetic terms that contribute to hot spot interactions, i.e. van der Waals potentials, solvation energy, hydrogen bonds and Coulomb electrostatics. We treat them as input features and use machine learning algorithms such as Support Vector Machines and Gaussian Processes to optimally combine and integrate them, based on a set of training examples of alanine mutations. We show that our approach is effective in predicting hot spots and it compares favourably to other available methods. In particular we find the best performances using Transductive Support Vector Machines, a semi-supervised learning scheme. When hot spots are defined as those residues for which DeltaDeltaG >or= 2 kcal/mol, our method achieves a precision and a recall respectively of 56% and 65%.

CONCLUSION

We have developed an hybrid scheme in which energy terms are used as input features of machine learning models. This strategy combines the strengths of machine learning and energy-based methods. Although so far these two types of approaches have mainly been applied separately to biomolecular problems, the results of our investigation indicate that there are substantial benefits to be gained by their integration.

Identifying native-like protein structures with scoring functions based on all-atom ECEPP force fields, implicit solvent models and structure relaxation

Availability of energy functions which can discriminate native-like from non-native protein conformations is crucial for theoretical protein structure prediction and refinement of low-resolution protein models. This article reports the results of benchmark tests for scoring functions based on two all-atom ECEPP force fields, that is, ECEPP/3 and ECEPP05, and two implicit solvent models for a large set of protein decoys. The following three scoring functions are considered: (i) ECEPP05 plus a solvent-accessible surface area model with the parameters optimized with a set of protein decoys (ECEPP05/SA); (ii) ECEPP/3 plus the solvent-accessible surface area model of Ooi et al. (Proc Natl Acad Sci USA 1987;84:3086-3090) (ECEPP3/OONS); and (iii) ECEPP05 plus an implicit solvent model based on a solution of the Poisson equation with an optimized Fast Adaptive Multigrid Boundary Element (FAMBEpH) method (ECEPP05/FAMBEpH). Short Monte Carlo-with-Minimization (MCM) simulations, following local energy minimization, are used as a scoring method with ECEPP05/SA and ECEPP3/OONS potentials, whereas energy calculation is used with ECEPP05/FAMBEpH. The performance of each scoring function is evaluated by examining its ability to distinguish between native-like and non-native protein structures. The results of the tests show that the new ECEPP05/SA scoring function represents a significant improvement over the earlier ECEPP3/OONS version of the force field. Thus, it is able to rank native-like structures with C(alpha) root-mean-square-deviations below 3.5 A as lowest-energy conformations for 76% and within the top 10 for 87% of the proteins tested, compared with 69 and 80%, respectively, for ECEPP3/OONS. The use of the FAMBEpH solvation model, which provides a more accurate description of the protein-solvent interactions, improves the discriminative ability of the scoring function to 89%. All failed tests in which the native-like structures cannot be discriminated as those with low energy, are due to omission of protein-protein interactions. The results of this study represent a benchmark in force-field development, and may be useful for evaluation of the performance of different force fields.

How well can the accuracy of comparative protein structure models be predicted?

Comparative structure models are available for two orders of magnitude more protein sequences than are experimentally determined structures. These models, however, suffer from two limitations that experimentally determined structures do not: They frequently contain significant errors, and their accuracy cannot be readily assessed. We have addressed the latter limitation by developing a protocol optimized specifically for predicting the Calpha root-mean-squared deviation (RMSD) and native overlap (NO3.5A) errors of a model in the absence of its native structure. In contrast to most traditional assessment scores that merely predict one model is more accurate than others, this approach quantifies the error in an absolute sense, thus helping to determine whether or not the model is suitable for intended applications. The assessment relies on a model-specific scoring function constructed by a support vector machine. This regression optimizes the weights of up to nine features, including various sequence similarity measures and statistical potentials, extracted from a tailored training set of models unique to the model being assessed: If possible, we use similarly sized models with the same fold; otherwise, we use similarly sized models with the same secondary structure composition. This protocol predicts the RMSD and NO3.5A errors for a diverse set of 580,317 comparative models of 6174 sequences with correlation coefficients (r) of 0.84 and 0.86, respectively, to the actual errors. This scoring function achieves the best correlation compared to 13 other tested assessment criteria that achieved correlations ranging from 0.35 to 0.71.

Use of decoys to optimize an all-atom force field including hydration

A novel method of parameter optimization is proposed. It makes use of large sets of decoys generated for six nonhomologous proteins with different architecture. Parameter optimization is achieved by creating a free energy gap between sets of nativelike and nonnative conformations. The method is applied to optimize the parameters of a physics-based scoring function consisting of the all-atom ECEPP05 force field coupled with an implicit solvent model (a solvent-accessible surface area model). The optimized force field is able to discriminate near-native from nonnative conformations of the six training proteins when used either for local energy minimization or for short Monte Carlo simulated annealing runs after local energy minimization. The resulting force field is validated with an independent set of six nonhomologous proteins, and appears to be transferable to proteins not included in the optimization; i.e., for five out of the six test proteins, decoys with 1.7- to 4.0-A all-heavy-atom root mean-square deviations emerge as those with the lowest energy. In addition, we examined the set of misfolded structures created by Park and Levitt using a four-state reduced model. The results from these additional calculations confirm the good discriminative ability of the optimized force field obtained with our decoy sets.

Divergence, recombination and retention of functionality during protein evolution

We have only a vague idea of precisely how protein sequences evolve in the context of protein structure and function. This is primarily because structural and functional contexts are not easily predictable from the primary sequence, and evaluating patterns of evolution at individual residue positions is also difficult. As a result of increasing biodiversity in genomics studies, progress is being made in detecting context-dependent variation in substitution processes, but it remains unclear exactly what context-dependent patterns we should be looking for. To address this, we have been simulating protein evolution in the context of structure and function using lattice models of proteins and ligands (or substrates). These simulations include thermodynamic features of protein stability and population dynamics. We refer to this approach as 'ab initio evolution' to emphasise the fact that the equilibrium details of fitness distributions arise from the physical principles of the system and not from any preconceived notions or arbitrary mathematical distributions. Here, we present results on the retention of functionality in homologous recombinants following population divergence. A central result is that protein structure characteristics can strongly influence recombinant functionality. Exceptional structures with many sequence options evolve quickly and tend to retain functionality--even in highly diverged recombinants. By contrast, the more common structures with fewer sequence options evolve more slowly, but the fitness of recombinants drops off rapidly as homologous proteins diverge. These results have implications for understanding viral evolution, speciation and directed evolutionary experiments. Our analysis of the divergence process can also guide improved methods for accurately approximating folding probabilities in more complex but realistic systems.

A composite score for predicting errors in protein structure models

Reliable prediction of model accuracy is an important unsolved problem in protein structure modeling. To address this problem, we studied 24 individual assessment scores, including physics-based energy functions, statistical potentials, and machine learning-based scoring functions. Individual scores were also used to construct approximately 85,000 composite scoring functions using support vector machine (SVM) regression. The scores were tested for their abilities to identify the most native-like models from a set of 6000 comparative models of 20 representative protein structures. Each of the 20 targets was modeled using a template of <30% sequence identity, corresponding to challenging comparative modeling cases. The best SVM score outperformed all individual scores by decreasing the average RMSD difference between the model identified as the best of the set and the model with the lowest RMSD (DeltaRMSD) from 0.63 A to 0.45 A, while having a higher Pearson correlation coefficient to RMSD (r=0.87) than any other tested score. The most accurate score is based on a combination of the DOPE non-hydrogen atom statistical potential; surface, contact, and combined statistical potentials from MODPIPE; and two PSIPRED/DSSP scores. It was implemented in the SVMod program, which can now be applied to select the final model in various modeling problems, including fold assignment, target-template alignment, and loop modeling.

Optimization of the GB/SA solvation model for predicting the structure of surface loops in proteins

Implicit solvation models are commonly optimized with respect to experimental data or Poisson-Boltzmann (PB) results obtained for small molecules, where the force field is sometimes not considered. In previous studies, we have developed an optimization procedure for cyclic peptides and surface loops in proteins based on the entire system studied and the specific force field used. Thus, the loop has been modeled by the simplified solvation function E(tot) = E(FF) (epsilon = 2r) + Sigma(i) sigma(i)A(i), where E(FF) (epsilon = nr) is the AMBER force field energy with a distance-dependent dielectric function, epsilon = nr, A(i) is the solvent accessible surface area of atom i, and sigma(i) is its atomic solvation parameter. During the optimization process, the loop is free to move while the protein template is held fixed in its X-ray structure. To improve on the results of this model, in the present work we apply our optimization procedure to the physically more rigorous solvation model, the generalized Born with surface area (GB/SA) (together with the all-atom AMBER force field) as suggested by Still and co-workers (J. Phys. Chem. A 1997, 101, 3005). The six parameters of the GB/SA model, namely, P(1)-P(5) and the surface area parameter, sigma (programmed in the TINKER package) are reoptimized for a "training" group of nine loops, and a best-fit set is defined from the individual sets of optimized parameters. The best-fit set and Still's original set of parameters (where Lys, Arg, His, Glu, and Asp are charged or neutralized) were applied to the training group as well as to a "test" group of seven loops, and the energy gaps and the corresponding RMSD values were calculated. These GB/SA results based on the three sets of parameters have been found to be comparable; surprisingly, however, they are somewhat inferior (e.g, of larger energy gaps) to those obtained previously from the simplified model described above. We discuss recent results for loops obtained by other solvation models and potential directions for future studies.

Free energy surfaces of miniproteins with a ββα motif: Replica exchange molecular dynamics simulation with an implicit solvation model

Designed miniproteins with a betabetaalpha motif, such as BBA5, 1FSD, and 1PSV can serve as a benchmark set to test the validity of all-atom force fields with computer simulation, because they contain all the basic structural elements in protein folding. Unfortunately, it was found that the standard all-atom force fields with the generalized Born (GB) implicit solvation model tend to produce distorted free energy surfaces for the betabetaalpha proteins, not only because energetically those proteins need to be described by more balanced weights of the alpha- and beta-strands, but also because the GB implicit solvation model suffers from overestimated salt bridge effects. In an attempt to resolve these problems, we have modified one of the standard all-atom force fields in conjunction with the GB model, such that each native state of the betabetaalpha proteins is in its free energy minimum state with reasonable energy barriers separating local minima. With this modified energy model, the free energy contour map in each protein was constructed from the replica exchange molecular dynamics REMD simulation. The resulting free energy surfaces are significantly improved in comparison with previous simulation results and consistent with general views on small protein folding behaviors with realistic topology and energetics of all three proteins.

Decoys for docking

Molecular docking is widely used to predict novel lead compounds for drug discovery. Success depends on the quality of the docking scoring function, among other factors. An imperfect scoring function can mislead by predicting incorrect ligand geometries or by selecting nonbinding molecules over true ligands. These false-positive hits may be considered "decoys". Although these decoys are frustrating, they potentially provide important tests for a docking algorithm; the more subtle the decoy, the more rigorous the test. Indeed, decoy databases have been used to improve protein structure prediction algorithms and protein-protein docking algorithms. Here, we describe 20 geometric decoys in five enzymes and 166 "hit list" decoys-i.e., molecules predicted to bind by our docking program that were tested and found not to do so-for beta-lactamase and two cavity sites in lysozyme. Especially in the cavity sites, which are very simple, these decoys highlight particular weaknesses in our scoring function. We also consider the performance of five other widely used docking scoring functions against our geometric and hit list decoys. Intriguingly, whereas many of these other scoring functions performed better on the geometric decoys, they typically performed worse on the hit list decoys, often highly ranking molecules that seemed to poorly complement the model sites. Several of these "hits"from the other scoring functions were tested experimentally and found, in fact, to be decoys. Collectively, these decoys provide a tool for the development and improvement of molecular docking scoring functions. Such improvements may, in turn, be rapidly tested experimentally against these and related experimental systems, which are well-behaved in assays and for structure determination.

Protein decoy sets for evaluating energy functions

Energy functions are crucial ingredients of protein tertiary structure prediction methods. Assessing the quality of energy functions is therefore of prime importance. It requires the elaboration of a standard evaluation scheme, whose key elements are: i). sets that contain the native and several non-native structures of proteins (decoys) in order to test whether the energy functions display the expected quality features and ii). measures to evaluate the reliability of energy functions. We present here a survey of the recent advances in these two related fields. In a first part, we analyze and review the large number of decoy sets that are available on the web, and we summarize the characteristics of a challenging decoy set. We then discuss how to define the quality of energy functions and review the measures related to it.

Domain swapping of CD4 upon dimerization

It has recently been shown that disulfide bond Cys130-Cys159 in domain 2 of monomeric CD4 is involved in the formation of CD4 disulfide-bonded dimers on cell surfaces and that it can influence the permissiveness of cells to HIV infection. Because this disulfide bond is buried in the monomer, a large conformational change must take place in order to allow for such disulfide exchange. Using standard optimization techniques, whose efficiency was first checked in the well-documented CD2 case, we have shown that 3D domain swapping is a likely candidate for the conformational change, the hinge loop, or linker, being loop E-F. Indeed, as a consequence of domain swapping, because Cys130 and Cys159 belong to beta-strands C and F, respectively, two disulfide bonds become established between Cys130 in one monomer and Cys159 in the other one. Such a disulfide exchange has already been observed when the nuclear magnetic resonance (NMR) structure of the prion protein was compared to the crystallographic, dimeric one. In both cases, domain swapping implies disulfide exchange because the linker is located in the sequence between two disulfide-bonded cysteines. As in the CD2 case, the proposed configuration of the CD4 dimer is found as a pair of neighboring monomers in the crystallographic unit cell. Moreover, because in this configuration the epitope of monoclonal antibody MT151, which does not compete with Gp120 for CD4 binding, is in the cleft between the pair of CD4 monomers, it is suggested that MT151 achieves its HIV-blocking activity by interfering with the formation of CD4 domain-swapped dimers on cell surface.