Recognition of distantly related proteins through energy calculations

Role of in silico structural modeling in predicting immunogenic neoepitopes for cancer vaccine development

In prior studies, we delineated the landscape of neoantigens arising from nonsynonymous point mutations in a murine pancreatic cancer model, Panc02. We developed a peptide vaccine by targeting neoantigens predicted using a pipeline that incorporates the MHC binding algorithm NetMHC. The vaccine, when combined with immune checkpoint modulators, elicited a robust neoepitope-specific antitumor immune response and led to tumor clearance. However, only a small fraction of the predicted neoepitopes induced T cell immunity, similarly to that reported for neoantigen vaccines tested in clinical studies. While these studies have used binding affinities to MHC I as surrogates for T cell immunity, this approach does not include spatial information on the mutated residue that is crucial for TCR activation. Here, we investigate conformational alterations in and around the MHC binding groove induced by selected minimal neoepitopes, and we examine the influence of a given mutated residue as a function of its spatial position. We found that structural parameters, including the solvent-accessible surface area (SASA) of the neoepitope and the position and spatial configuration of the mutated residue within the sequence, can be used to improve the prediction of immunogenic neoepitopes for inclusion in cancer vaccines.

Structural parameters, including the solvent exposed surface area of the neoepitope and the position and spatial configuration of the mutated residue can be used to improve the prediction of immunogenic neoepitopes for inclusion in cancer vaccines.

Discovery of holoenzyme-disrupting chemicals as substrate-selective CK2 inhibitors

CK2 is a constitutively active protein kinase overexpressed in numerous malignancies. Interaction between CK2α and CK2β subunits is essential for substrate selectivity. The CK2α/CK2β interface has been previously targeted by peptides to achieve functional effects; however, no small molecules modulators were identified due to pocket flexibility and open shape. Here we generated numerous plausible conformations of the interface using the fumigation modeling protocol, and virtually screened a compound library to discover compound 1 that suppressed CK2α/CK2β interaction in vitro and inhibited CK2 in a substrate-selective manner. Orthogonal SPR, crystallography, and NMR experiments demonstrated that 4 and 6, improved analogs of 1, bind to CK2α as predicted. Both inhibitors alter CK2 activity in cells through inhibition of CK2 holoenzyme formation. Treatment with 6 suppressed MDA-MB231 triple negative breast cancer cell growth and induced apoptosis. Altogether, our findings exemplify an innovative computational-experimental approach and identify novel non-peptidic inhibitors of CK2 subunit interface disclosing substrate-selective functional effects.

The structure function of the death domain of human IRAK-M

BACKGROUND

IRAK-M is an inhibitor of Toll-like receptor signaling that acts by re-directing IRAK-4 activity to TAK1 independent NF-κB activation and by inhibition of IRAK-1/IRAK-2 activity. IRAK-M is expressed in monocytes/macrophages and lung epithelial cells. Lack of IRAK-M in mice greatly improves the resistance to nosocomial pneumonia and lung tumors, which entices IRAK-M as a potential therapeutic target. IRAK-M consists of an N-terminal death domain (DD), a dysfunctional kinase domain and unstructured C-terminal domain. Little is known however on IRAK-M's structure-function relationships.

RESULTS

Since death domains provide the important interactions of IRAK-1, IRAK-2 and IRAK-4 molecules, we generated a 3D structure model of the human IRAK-M-DD (residues C5-G119) to guide mutagenesis studies and predict protein-protein interaction points. First we identified the DD residues involved in the endogenous capacity of IRAK-M to activate NF-κB that is displayed upon overexpression in 293T cells. W74 and R97, at distinct interfaces of the IRAK-M-DD, were crucial for this endogenous NF-κB activating capacity, as well as the C-terminal domain (S445-E596) of IRAK-M. Resulting anti-inflammatory A20 and pro-inflammatory IL-8 transcription in 293T cells was W74 dependent, while IL-8 protein expression was dependent on R97 and the TRAF6 binding motif at P478. The IRAK-M-DD W74 and R97 binding interfaces are predicted to interact with opposite sides of IRAK-4-DD's. Secondly we identified DD residues important for the inhibitory action of IRAK-M by stable overexpression of mutants in THP-1 macrophages and H292 lung epithelial cells. IRAK-M inhibited TLR2/4-mediated cytokine production in macrophages in a manner that is largely dependent on W74. R97 was not involved in inhibition of TNF production but was engaged in IL-6 down-regulation by IRAK-M. Protein-interactive residues D19-A23, located in between W74 and R97, were also observed to be crucial for inhibition of TLR2/4 mediated cytokine induction in macrophages. Remarkably, IRAK-M inhibited TLR5 mediated IL-8 production by lung epithelial cells independent of W74 and R97, but dependent on D19-A23 and R70, two surface-exposed regions that harbor predicted IRAK-2-DD interaction points of IRAK-M.

CONCLUSION

IRAK-M employs alternate residues of its DD to inhibit the different inflammatory mediators induced by varying TLRs and cells.

Discovery of novel membrane binding structures and functions

The function of a protein is determined by its intrinsic activity in the context of its subcellular distribution. Membranes localize proteins within cellular compartments and govern their specific activities. Discovering such membrane-protein interactions is important for understanding biological mechanisms and could uncover novel sites for therapeutic intervention. We present a method for detecting membrane interactive proteins and their exposed residues that insert into lipid bilayers. Although the development process involved analysis of how C1b, C2, ENTH, FYVE, Gla, pleckstrin homology (PH), and PX domains bind membranes, the resulting membrane optimal docking area (MODA) method yields predictions for a given protein of known three-dimensional structures without referring to canonical membrane-targeting modules. This approach was tested on the Arf1 GTPase, ATF2 acetyltransferase, von Willebrand factor A3 domain, and Neisseria gonorrhoeae MsrB protein and further refined with membrane interactive and non-interactive FAPP1 and PKD1 pleckstrin homology domains, respectively. Furthermore we demonstrate how this tool can be used to discover unprecedented membrane binding functions as illustrated by the Bro1 domain of Alix, which was revealed to recognize lysobisphosphatidic acid (LBPA). Validation of novel membrane-protein interactions relies on other techniques such as nuclear magnetic resonance spectroscopy (NMR), which was used here to map the sites of micelle interaction. Together this indicates that genome-wide identification of known and novel membrane interactive proteins and sites is now feasible and provides a new tool for functional annotation of the proteome.

Systematic exploitation of multiple receptor conformations for virtual ligand screening

The role of virtual ligand screening in modern drug discovery is to mine large chemical collections and to prioritize for experimental testing a comparatively small and diverse set of compounds with expected activity against a target. Several studies have pointed out that the performance of virtual ligand screening can be improved by taking into account receptor flexibility. Here, we systematically assess how multiple crystallographic receptor conformations, a powerful way of discretely representing protein plasticity, can be exploited in screening protocols to separate binders from non-binders. Our analyses encompass 36 targets of pharmaceutical relevance and are based on actual molecules with reported activity against those targets. The results suggest that an ensemble receptor-based protocol displays a stronger discriminating power between active and inactive molecules as compared to its standard single rigid receptor counterpart. Moreover, such a protocol can be engineered not only to enrich a higher number of active compounds, but also to enhance their chemical diversity. Finally, some clear indications can be gathered on how to select a subset of receptor conformations that is most likely to provide the best performance in a real life scenario.

Four-dimensional docking: a fast and accurate account of discrete receptor flexibility in ligand docking

Many available methods aimed at incorporating the receptor flexibility in ligand docking are computationally expensive, require a high level of user intervention, and were tested only on benchmarks of limited size and diversity. Here we describe the four-dimensional (4D) docking approach that allows seamless incorporation of receptor conformational ensembles in a single docking simulation and reduces the sampling time while preserving the accuracy of traditional ensemble docking. The approach was tested on a benchmark of 99 therapeutically relevant proteins and 300 diverse ligands (half of them experimental or marketed drugs). The conformational variability of the binding pockets was represented by the available crystallographic data, with the total of 1113 receptor structures. The 4D docking method reproduced the correct ligand binding geometry in 77.3% of the benchmark cases, matching the success rate of the traditional approach but employed on average only one-fourth of the time during the ligand sampling phase.

A new method for ligand docking to flexible receptors by dual alanine scanning and refinement (SCARE)

Protein binding sites undergo ligand specific conformational changes upon ligand binding. However, most docking protocols rely on a fixed conformation of the receptor, or on the prior knowledge of multiple conformations representing the variation of the pocket, or on a known bounding box for the ligand. Here we described a general induced fit docking protocol that requires only one initial pocket conformation and identifies most of the correct ligand positions as the lowest score. We expanded a previously used diverse "cross-docking" benchmark to thirty ligand-protein pairs extracted from different crystal structures. The algorithm systematically scans pairs of neighbouring side chains, replaces them by alanines, and docks the ligand to each 'gapped' version of the pocket. All docked positions are scored, refined with original side chains and flexible backbone and re-scored. In the optimal version of the protocol pairs of residues were replaced by alanines and only one best scoring conformation was selected from each 'gapped' pocket for refinement. The optimal SCARE (SCan Alanines and REfine) protocol identifies a near native conformation (under 2 angstroms RMSD) as the lowest rank for 80% of pairs if the docking bounding box is defined by the predicted pocket envelope, and for as many as 90% of the pairs if the bounding box is derived from the known answer with approximately 5 angstroms margin as used in most previous publications. The presented fully automated algorithm takes about 2 h per pose of a single processor time, requires only one pocket structure and no prior knowledge about the binding site location. Furthermore, the results for conformationally conserved pockets do not deteriorate due to substantial increase of the pocket variability.

Degradation of the amyloid beta-protein by the novel mitochondrial peptidasome, PreP

Recently we have identified the novel mitochondrial peptidase responsible for degrading presequences and other short unstructured peptides in mitochondria, the presequence peptidase, which we named PreP peptidasome. In the present study we have identified and characterized the human PreP homologue, hPreP, in brain mitochondria, and we show its capacity to degrade the amyloid beta-protein (Abeta). PreP belongs to the pitrilysin oligopeptidase family M16C containing an inverted zinc-binding motif. We show that hPreP is localized to the mitochondrial matrix. In situ immuno-inactivation studies in human brain mitochondria using anti-hPreP antibodies showed complete inhibition of proteolytic activity against Abeta. We have cloned, overexpressed, and purified recombinant hPreP and its mutant with catalytic base Glu(78) in the inverted zinc-binding motif replaced by Gln. In vitro studies using recombinant hPreP and liquid chromatography nanospray tandem mass spectrometry revealed novel cleavage specificities against Abeta-(1-42), Abeta-(1-40), and Abeta Arctic, a protein that causes increased protofibril formation an early onset familial variant of Alzheimer disease. In contrast to insulin degrading enzyme, which is a functional analogue of hPreP, hPreP does not degrade insulin but does degrade insulin B-chain. Molecular modeling of hPreP based on the crystal structure at 2.1 A resolution of AtPreP allowed us to identify Cys(90) and Cys(527) that form disulfide bridges under oxidized conditions and might be involved in redox regulation of the enzyme. Degradation of the mitochondrial Abeta by hPreP may potentially be of importance in the pathology of Alzheimer disease.

Fold recognition by combining sequence profiles derived from evolution and from depth-dependent structural alignment of fragments

Recognizing structural similarity without significant sequence identity has proved to be a challenging task. Sequence-based and structure-based methods as well as their combinations have been developed. Here, we propose a fold-recognition method that incorporates structural information without the need of sequence-to-structure threading. This is accomplished by generating sequence profiles from protein structural fragments. The structure-derived sequence profiles allow a simple integration with evolution-derived sequence profiles and secondary-structural information for an optimized alignment by efficient dynamic programming. The resulting method (called SP(3)) is found to make a statistically significant improvement in both sensitivity of fold recognition and accuracy of alignment over the method based on evolution-derived sequence profiles alone (SP) and the method based on evolution-derived sequence profile and secondary structure profile (SP(2)). SP(3) was tested in SALIGN benchmark for alignment accuracy and Lindahl, PROSPECTOR 3.0, and LiveBench 8.0 benchmarks for remote-homology detection and model accuracy. SP(3) is found to be the most sensitive and accurate single-method server in all benchmarks tested where other methods are available for comparison (although its results are statistically indistinguishable from the next best in some cases and the comparison is subjected to the limitation of time-dependent sequence and/or structural library used by different methods.). In LiveBench 8.0, its accuracy rivals some of the consensus methods such as ShotGun-INBGU, Pmodeller3, Pcons4, and ROBETTA. SP(3) fold-recognition server is available on http://theory.med.buffalo.edu.

SPARKS 2 and SP3 servers in CASP6

Two single-method servers, SPARKS 2 and SP3, participated in automatic-server predictions in CASP6. The overall results for all as well as detailed performance in comparative modeling targets are presented. It is shown that both SPARKS 2 and SP3 are able to recognize their corresponding best templates for all easy comparative modeling targets. The alignment accuracy, however, is not always the best among all the servers. Possible factors are discussed. SPARKS 2 and SP3 fold recognition servers, as well as their executables, are freely available for all academic users on http://theory.med.buffalo.edu.

A machine learning information retrieval approach to protein fold recognition

MOTIVATION

Recognizing proteins that have similar tertiary structure is the key step of template-based protein structure prediction methods. Traditionally, a variety of alignment methods are used to identify similar folds, based on sequence similarity and sequence-structure compatibility. Although these methods are complementary, their integration has not been thoroughly exploited. Statistical machine learning methods provide tools for integrating multiple features, but so far these methods have been used primarily for protein and fold classification, rather than addressing the retrieval problem of fold recognition-finding a proper template for a given query protein.

RESULTS

Here we present a two-stage machine learning, information retrieval, approach to fold recognition. First, we use alignment methods to derive pairwise similarity features for query-template protein pairs. We also use global profile-profile alignments in combination with predicted secondary structure, relative solvent accessibility, contact map and beta-strand pairing to extract pairwise structural compatibility features. Second, we apply support vector machines to these features to predict the structural relevance (i.e. in the same fold or not) of the query-template pairs. For each query, the continuous relevance scores are used to rank the templates. The FOLDpro approach is modular, scalable and effective. Compared with 11 other fold recognition methods, FOLDpro yields the best results in almost all standard categories on a comprehensive benchmark dataset. Using predictions of the top-ranked template, the sensitivity is approximately 85, 56, and 27% at the family, superfamily and fold levels respectively. Using the 5 top-ranked templates, the sensitivity increases to 90, 70, and 48%.

Single-body residue-level knowledge-based energy score combined with sequence-profile and secondary structure information for fold recognition

An elaborate knowledge-based energy function is designed for fold recognition. It is a residue-level single-body potential so that highly efficient dynamic programming method can be used for alignment optimization. It contains a backbone torsion term, a buried surface term, and a contact-energy term. The energy score combined with sequence profile and secondary structure information leads to an algorithm called SPARKS (Sequence, secondary structure Profiles and Residue-level Knowledge-based energy Score) for fold recognition. Compared with the popular PSI-BLAST, SPARKS is 21% more accurate in sequence-sequence alignment in ProSup benchmark and 10%, 25%, and 20% more sensitive in detecting the family, superfamily, fold similarities in the Lindahl benchmark, respectively. Moreover, it is one of the best methods for sensitivity (the number of correctly recognized proteins), alignment accuracy (based on the MaxSub score), and specificity (the average number of correctly recognized proteins whose scores are higher than the first false positives) in LiveBench 7 among more than twenty servers of non-consensus methods. The simple algorithm used in SPARKS has the potential for further improvement. This highly efficient method can be used for fold recognition on genomic scales. A web server is established for academic users on http://theory.med.buffalo.edu.

Enzyme family classification by support vector machines

One approach for facilitating protein function prediction is to classify proteins into functional families. Recent studies on the classification of G-protein coupled receptors and other proteins suggest that a statistical learning method, Support vector machines (SVM), may be potentially useful for protein classification into functional families. In this work, SVM is applied and tested on the classification of enzymes into functional families defined by the Enzyme Nomenclature Committee of IUBMB. SVM classification system for each family is trained from representative enzymes of that family and seed proteins of Pfam curated protein families. The classification accuracy for enzymes from 46 families and for non-enzymes is in the range of 50.0% to 95.7% and 79.0% to 100% respectively. The corresponding Matthews correlation coefficient is in the range of 54.1% to 96.1%. Moreover, 80.3% of the 8,291 correctly classified enzymes are uniquely classified into a specific enzyme family by using a scoring function, indicating that SVM may have certain level of unique prediction capability. Testing results also suggest that SVM in some cases is capable of classification of distantly related enzymes and homologous enzymes of different functions. Effort is being made to use a more comprehensive set of enzymes as training sets and to incorporate multi-class SVM classification systems to further enhance the unique prediction accuracy. Our results suggest the potential of SVM for enzyme family classification and for facilitating protein function prediction. Our software is accessible at http://jing.cz3.nus.edu.sg/cgi-bin/svmprot.cgi.

ICFF: a new method to incorporate implicit flexibility into an internal coordinate force field

We introduce a new method to accurately "project" a Cartesian force field onto an internal coordinate molecular model with fixed-bond geometry. The algorithm automatically generates the Internal Coordinate Force Field (ICFF), which is a close approximation of the "source" Cartesian force field. The ICFF method reduces the number of free variables in a model by at least 10-fold and facilitates the fast convergence of geometry optimizations, an advantage that is critical for many applications such as the docking of flexible ligands or conformational modeling of macromolecules. Although covalent geometry is fixed in an ICFF model, implicit flexibility is incorporated into the force field parameters in the following two ways. First, we formulate an empirical torsion energy term in ICFF as a sixfold Fourier series and develop a procedure to calculate the Fourier coefficients from the conformational energy profiles of the fully flexible Cartesian model. The ICFF torsion parameters thus represent not only torsion component of the source force field, but also bond bending, bond stretching, and "1-4" van der Waals interactions. Second, we use a soft polynomial repulsion function for "1-5" and "1-6" interactions to mimic the flexibility of bonds, connecting these atoms. Also, we suggest a way to use a local part of the Cartesian force field to automatically generate fixed covalent geometries, compatible with the ICFF energy function. Here, we present an implementation of the ICFF algorithm, which employs the MMFF94s Cartesian force field as a "source." Extensive benchmarking of ICFF with a representative set of organic molecules demonstrates that the implicit flexibility model accurately reproduces MMFF94s equilibrium conformational energy differences (RMSD approximately 0.64 kcal) and, most importantly, detailed torsion energy profiles (RMSD approximately 0.37 kcal). This accuracy is characteristic of the method, because all the ICFF parameters (except one scaling factor in the "1-5,1-6" repulsion term) are derived directly from the source Cartesian force field and do not depend on any particular molecular set. In contrast, the rigid geometry model with the MMFF94s energy function yields highly biased estimations in this test with the RMSD exceeding 1.2 kcal for the equilibrium energy comparisons and approximately 3.4 kcal for the torsion energy profiles.

Completion and refinement of 3-D homology models with restricted molecular dynamics: application to targets 47, 58, and 111 in the CASP modeling competition and posterior analysis

A method is presented to refine models built by homology by the use of restricted molecular dynamics (MD) techniques. The basic idea behind this method is the use of structure validation software to determine for each residue the likelihood that it is modeled correctly. This information is used to determine constraints and restraints in an MD simulation including explicit solvent molecules, which is used for model refinement. The procedure is based on the idea that residues that the validation software identifies as correctly positioned should be strongly constrained or restrained in the MD simulations, whereas residues that are likely to be positioned wrongly should move freely. Two different protocols are compared: one (applied to CASP3 target T58) using full structural constraints with separate optimization of each short fragment and the other (applied to T47) allowing some freedom using harmonic restraining potentials, with automatic optimization of the whole molecule. Structures along the MD trajectory that scored best in structural checks were selected for the construction of models that appeared to be successful in the CASP3 competition. Model refinement with MD in general leads to a model that is less like the experimental structure (Levitt et al. Nature Struct Biol 1999;6:108-111). Actually, refined T47 was slightly improved compared to the starting model; changes in model T58 led not to further enhancement. After the X-ray structure of the modeled proteins became known, the procedure was evaluated for two targets (T47 and the CASP4 target T111) by comparing a long simulation in water with the experimental target structures. It was found that structural improvements could be obtained on a nanosecond time scale by allowing appropriate freedom in the simulation. Structural checks applied to fast fluctuations do not appear to be informative for the correctness of the structure. However, both a simple hydrogen bond count and a simple compactness measure, if averaged over times of typically 300 ps, correlate well with structural correctness and we suggest that criteria based on these properties may be used in computational folding strategies.

Studies in the assessment of folding quality for protein modeling and structure prediction

A diagnostic for assessing the quality of a fold has been developed to which further criteria can be progressively added. The goal is to create a measure that can follow the status of a protein structure in a simulation or modeling process, when the answer (the experimental structure) is not known in advance, rather than simply reject deliberate misfolds. This places greater emphasis on the need to study, and calibrate against, marginal cases, i.e., unusual native structures, incomplete structures, partially erroneous X-ray structures, good models, poor models, and the effect of cofactors. The first three terms introduced in the diagnostic are appropriate core-forming properties or noncore properties of residues in relation to tertiary structure, appropriate neighboring structure density for each residue in relation to tertiary structure, and secondary structure consistency. While the method emerges as a useful simulation analysis tool, we find a need for further fine-tuning to diminish sensitivity to minor conformational changes that retain essential features of the fold, balanced against the need to obtain a more sensitive response when a conformational change involves less physically meaningful interatomic interactions. This dual utility is difficult to obtain: the investigation highlights some of the issues. Initial attempts to obtain it have led to terms in the diagnostic that are admittedly complex: simplifications must also be explored.

Genome-based structural biology

Spectacular achievements in whole genome sequencing open up new possibilities for structural research. Protein structures can now be studied in their natural genomic context. On the other hand, structure prediction algorithms can be improved using species-specific tendencies in folding patterns. Finally, efficient strategies to select targets for structure determination can be devised. In this review we consider new computational approaches and results in protein structure analysis stemming from the availability of complete genomes.

GenTHREADER: an efficient and reliable protein fold recognition method for genomic sequences

A new protein fold recognition method is described which is both fast and reliable. The method uses a traditional sequence alignment algorithm to generate alignments which are then evaluated by a method derived from threading techniques. As a final step, each threaded model is evaluated by a neural network in order to produce a single measure of confidence in the proposed prediction. The speed of the method, along with its sensitivity and very low false-positive rate makes it ideal for automatically predicting the structure of all the proteins in a translated bacterial genome (proteome). The method has been applied to the genome of Mycoplasma genitalium, and analysis of the results shows that as many as 46 % of the proteins derived from the predicted protein coding regions have a significant relationship to a protein of known structure. In some cases, however, only one domain of the protein can be predicted, giving a total coverage of 30 % when calculated as a fraction of the number of amino acid residues in the whole proteome.

Energy strain in three-dimensional protein structures

BACKGROUND

Steric strain in protein three-dimensional structures is related to unfavorable inter-atomic interactions. The steric strain may be a result of packing or functional requirements, or may indicate an error in the coordinates of a structure. Detailed energy functions are, however, usually considered too noisy for error detection.

RESULTS

After a short energy refinement, a full-atom, detailed energy function becomes a sensitive indicator of errors. The statistics of the energy distribution of amino acid residues in high-resolution crystal structures, represented by models with idealized covalent geometry, were calculated. The interaction energy of each residue with the whole protein structure and with the solvent was considered. Normalized deviations of amino acid residue energies from their average values were used for detecting energy-strained and, therefore, potentially incorrect fragments of a polypeptide chain. Protein three-dimensional structures of different origin (X-ray crystallography, NMR spectroscopy, theoretical models and deliberately misfolded decoys) were compared. Examples of the applications to loop and homology modeling are provided.

CONCLUSIONS

Elevated levels of energy strain may point at a problematic fragment in a protein three-dimensional structure of either experimental or theoretical origin. The approach may be useful in model building and refinement, modeling by homology, protein design, folding calculations, and protein structure analysis.

Do aligned sequences share the same fold?

Sequence comparison remains a powerful tool to assess the structural relatedness of two proteins. To develop a sensitive sequence-based procedure for fold recognition, we performed an exhaustive global alignment (with zero end gap penalties) between sequences of protein domains with known three-dimensional folds. The subset of 1.3 million alignments between sequences of structurally unrelated domains was used to derive a set of analytical functions that represent the probability of structural significance for any sequence alignment at a given sequence identity, sequence similarity and alignment score. Analysis of overlap between structurally significant and insignificant alignments shows that sequence identity and sequence similarity measures are poor indicators of structural relatedness in the "twilight zone", while the alignment score allows much better discrimination between alignments of structurally related and unrelated sequences for a wide variety of alignment settings. A fold recognition benchmark was used to compare eight different substitution matrices with eight sets of gap penalties. The best performing matrices were Gonnet and Blosum50 with normalized gap penalties of 2.4/0.15 and 2.0/0.15, respectively, while the positive matrices were the worst performers. The derived functions and parameters can be used for fold recognition via a multilink chain of probability weighted pairwise sequence alignments.

Protein fold recognition by prediction-based threading

In fold recognition by threading one takes the amino acid sequence of a protein and evaluates how well it fits into one of the known three-dimensional (3D) protein structures. The quality of sequence-structure fit is typically evaluated using inter-residue potentials of mean force or other statistical parameters. Here, we present an alternative approach to evaluating sequence-structure fitness. Starting from the amino acid sequence we first predict secondary structure and solvent accessibility for each residue. We then thread the resulting one-dimensional (1D) profile of predicted structure assignments into each of the known 3D structures. The optimal threading for each sequence-structure pair is obtained using dynamic programming. The overall best sequence-structure pair constitutes the predicted 3D structure for the input sequence. The method is fine-tuned by adding information from direct sequence-sequence comparison and applying a series of empirical filters. Although the method relies on reduction of 3D information into 1D structure profiles, its accuracy is, surprisingly, not clearly inferior to methods based on evaluation of residue interactions in 3D. We therefore hypothesise that existing 1D-3D threading methods essentially do not capture more than the fitness of an amino acid sequence for a particular 1D succession of secondary structure segments and residue solvent accessibility. The prediction-based threading method on average finds any structurally homologous region at first rank in 29% of the cases (including sequence information). For the 22% first hits detected at highest scores, the expected accuracy rose to 75%. However, the task of detecting entire folds rather than homologous fragments was managed much better; 45 to 75% of the first hits correctly recognised the fold.

Progress in protein structure prediction

If protein structure prediction methods are to make any impact on the impending onerous task of analyzing the large numbers of unknown protein sequences generated by the ongoing genome-sequencing projects, it is vital that they make the difficult transition from computational 'gedankenexperiments' to practical software tools. This has already happened in the field of comparative modelling and is currently happening in the threading field. Unfortunately, there is little evidence of this transition happening in the field of ab initio tertiary-structure prediction.

Protein structure prediction force fields: parametrization with quasi-newtonian dynamics

We present an unusual method for parametrizing low-resolution force fields of the type used for protein structure prediction. Force field parameters were determined by assigning each a fictitious mass and using a quasi-molecular dynamics algorithm in parameter space. The quasi-energy term favored folded native structures and specifically penalized folded nonnative structures. The force field was generated after optimizing less than 70 adjustable parameters, but shows a strong ability to discriminate between native structures and compact misfolded alternatives. The functional form of the force field was chosen as in molecular mechanics and is not table-driven. It is continuous with continuous derivatives and is thus suitable for use with algorithms such as energy minimization or newtonian dynamics.

A new method for modeling large-scale rearrangements of protein domains

A method for modeling large-scale rearrangements of protein domains connected by a single- or a double-stranded linker is proposed. Multidomain proteins may undergo substantial domain displacements, while their intradomain structure remains essentially unchanged. The method allows automatic identification of an interdomain linker and builds an all-atom model of a protein structure in internal coordinates. Torsion angles belonging to the interdomain linkers and side chains potentially able to form domain interfaces are set free while remaining torsions, bond lengths, and bond angles are fixed. Large-scale sampling of the reduced torsions conformational subspace is effected with the "biased probability Monte Carlo-minimization" method [Abagyan, R.A., Totrov, M.M. (1994): J. Mol. Biol. 235, 983-1002]. Solvation and side-chain entropic contributions are added to the energy function. A special procedure has been developed to generate concerted deformations of a double-stranded interdomain linker in such a way that the polypeptide chain continuity is preserved. The method was tested on Bence-Jones protein with a single-stranded linker and lysine/arginine/ornithine-binding (LAO) protein with a double-stranded linker. For each protein, structurally diverse low-energy conformations with ideal covalent geometry were generated, and an overlap between two sets of conformations generated starting from the crystallographically determined "closed" and "open" forms was found. One of the low-energy conformations generated in a run starting from the LAO "closed" form was only 2.2 A away from the structure of the "open" form. The method can be useful in predicting the scope of possible domain rearrangements of a multidomain protein.

Protein structure: what is it possible to predict now?

The computational techniques of sorting out protein folds (these techniques include dynamic programming, self-consistent field theory, etc.) have already ceased to be the bottleneck of predictions. The main problem is that all the methods of recognition and prediction of protein structure can actually use only some part of the interactions operating in the chain, and that even their energies are not known precisely. This is the principal source of errors now. The errors can be reduced by employment of many distant homologues, but this opens a possibility to predict a generalized folding pattern rather than a particular fold with all its details.

Potential energy functions for threading

Despite little progress in ab initio solutions to the problem of predicting a protein's tertiary structure, over the past four years or so the development of fold-recognition methods for tertiary structure prediction has been the source of some encouragement in this difficult field. Despite promising initial results, these methods are clearly not yet fully mature and many groups are now working on different aspects of the methods involved in the hope of increasing the reliability and sensitivity of these tools.

A study of combined structure/sequence profiles

BACKGROUND

For genome sequencing projects to achieve their full impact on biology and medicine, each protein sequence must be identified with its three-dimensional structure. Fold assignment methods (also called profile and threading methods) attempt to assign sequences to known protein folds by computing the compatibility of sequence to fold.

RESULTS

We have extended profile methods for the detection of protein folds having structural similarity but low sequence similarity to sequence probes. Our extension combines sequence substitution tables with structural properties to form a combined profile. The structural properties used in this study include distances between residues, exposed areas, areas buried by polar atoms, and properties of the original three-dimensional profile method. We compared the performance of these combined profiles with different sequence matrices and with the original three-dimensional profile method. To determine the optimal gap penalties and weights used with these profiles, we employed a genetic algorithm. The performance of these combined profiles was tested by cross validation using independent test and training sets.

CONCLUSIONS

These studies show that the combined profiles perform better than profiles based on either structural or sequence information alone.

Homology modeling by the ICM method

Five models have been built by the ICM method for the Comparative Modeling section of the Meeting on the Critical Assessment of Techniques for Protein Structure Prediction. The targets have homologous proteins with known three-dimensional structure with sequence identity ranging from 25 to 77%. After alignment of the target sequence with the related three-dimensional structure, the modeling procedure consists of two subproblems: side-chain prediction and loop prediction. The ICM method approaches these problems with the following steps: (1) a starting model is created based on the homologous structure with the conserved portion fixed and the nonconserved portion having standard covalent geometry and free torsion angles; (2) the Biased Probability Monte Carlo (BPMC) procedure is applied to search the subspaces of either all the nonconservative side-chain torsion angles or torsion angles in a loop backbone and surrounding side chains. A special algorithm was designed to generate low-energy loop deformations. The BPMC procedure globally optimizes the energy function consisting of ECEPP/3 and solvation energy terms. Comparison of the predictions with the NMR or crystallographic solutions reveals a high proportion of correctly predicted side chains. The loops were not correctly predicted because imprinted distortions of the backbone increased the energy of the near-native conformation and thus made the solution unrecognizable. Interestingly, the energy terms were found to be reliable and the sampling of conformational space sufficient. The implications of this finding for the strategies of future comparative modeling are discussed.

Threading a database of protein cores

We present an analysis of 10 blind predictions prepared for a recent conference, "Critical Assessment of Techniques for Protein Structure Prediction." The sequences of these proteins are not detectably similar to those of any protein in the structure database then available, but we attempted, by a threading method, to recognize similarity to known domain folds. Four of the 10 proteins, as we subsequently learned, do indeed show significant similarity to then-known structures. For 2 of these proteins the predictions were accurate, in the sense that a similar structure was at or near the top of the list of threading scores, and the threading alignment agreed well with the corresponding structural alignment. For the best predicted model mean alignment error relative to the optimal structural alignment was 2.7 residues, arising entirely from small "register shifts" of strands or helices. In the analysis we attempt to identify factors responsible for these successes and failures. Since our threading method does not use gap penalties, we may readily distinguish between errors arising from our prior definition of the "cores" of known structures and errors arising from inherent limitations in the threading potential. It would appear from the results that successful substructure recognition depends most critically on accurate definition of the "fold" of a database protein. This definition must correctly delineate substructures that are, and are not, likely to be conserved during protein evolution.

Modeling mutations and homologous proteins

A protein sequence with at lease 40% identity to a known structure can now be modelled automatically, with an accuracy approaching that o fa low-resolution X-ray structure or a medium-resolution nuclear magnetic resonance structure. In general, these models have goods stereochemistry and an overall structural accuracy that is as high as the similarity between the template and the actual structure being predicted. As a result, the number of sequences that can be modelled is an order of magnitude larger then the number of experimentally determined protein structures. In addition, evaluation techniques are available that can estimated errors in different regions of the model. Thus, the number of applications where homology modelling is proving useful is growing rapidly.

Theoretical approaches to designing novel sequences to fit a given fold

One of the major goals of molecular biology is to understand how protein chains fold into a unique three-dimensional structure. Given this knowledge, perhaps the most exciting prospect will be the possibility of designing new proteins to perform designated tasks. The eventual pinnacle of protein engineering will be the fully automated design of a protein with novel structure and function. Achievement of this aim lies far in the future, although some early progress has been made recently.

Statistics of sequence-structure threading

The past two years have seen the rapid development of new recognition methods for protein structure prediction. These algorithms 'thread' the sequence of one protein through the known structure of another, looking for an alignment that corresponds to an energetically favorable model structure. Because they are based on energy calculation, rather than evolutionary distance, these methods extend the possibility of structure prediction by comparative modeling to a larger class of new sequences, where similarity to known structures is recognizable by no other means. The strength of the evidence they offer should be judged by objective statistical tests, however, so as to rule out the possibility that favorable scores arise from chance factors such as similarity of length, composition, or the consideration of a large number of alternative alignments. Calculation of objective p-values by analytical means is not yet possible, but it would appear that approximate values may be obtained by simulation, as they are in gapped, global sequence alignment. We propose that the results of threading experiments should include Z-scores relative to the composition-corrected score distribution obtained for shuffled and optimally aligned sequences.

Protein structure prediction: recognition of primary, secondary, and tertiary structural features from amino acid sequence

This review attempts a critical stock-taking of the current state of the science aimed at predicting structural features of proteins from their amino acid sequences. At the primary structure level, methods are considered for detection of remotely related sequences and for recognizing amino acid patterns to predict posttranslational modifications and binding sites. The techniques involving secondary structural features include prediction of secondary structure, membrane-spanning regions, and secondary structural class. At the tertiary structural level, methods for threading a sequence into a mainchain fold, homology modeling and assigning sequences to protein families with similar folds are discussed. A literature analysis suggests that, to date, threading techniques are not able to show their superiority over sequence pattern recognition methods. Recent progress in the state of ab initio structure calculation is reviewed in detail. The analysis shows that many structural features can be predicted from the amino acid sequence much better than just a few years ago and with attendant utility in experimental research. Best prediction can be achieved for new protein sequences that can be assigned to well-studied protein families. For single sequences without homologues, the folding problem has not yet been solved.

Sensitive methods for determining the relatedness of proteins with limited sequence homology

Recently, considerable advances have been made in attempts to determine the relatedness of protein sequences distant in evolution, when little or no knowledge is available concerning the corresponding tertiary architectures. Several improvements have been made to existing techniques, and these include better amino acid substitution weights contained in scoring matrices, better understanding of the effect of different gap penalty values in delineating the optimal alignment of two sequences, improved assessment of the significance of suggested sequence similarities, consideration of high scoring alternative alignments, and advances in searching entire sequence databases with the profile technique utilizing multiple-sequence information. New approaches that search for similarity a query sequence against large data banks rely on highly conserved segmental motifs defined from an aligned family of sequences. A sensitive algorithm to find distant repeats within one primary structure has also been developed recently. Solution of the inverse protein-folding problem, which involves an estimation of the ability of a sequence to take on a known main-chain tertiary topology (despite little homology with the known sequence), is being facilitated by the recent explosion in the number of new algorithms.