Assigning folds to the proteins encoded by the genome of Mycoplasma genitalium

Recent Progress of Protein Tertiary Structure Prediction

The prediction of three-dimensional (3D) protein structure from amino acid sequences has stood as a significant challenge in computational and structural bioinformatics for decades. Recently, the widespread integration of artificial intelligence (AI) algorithms has substantially expedited advancements in protein structure prediction, yielding numerous significant milestones. In particular, the end-to-end deep learning method AlphaFold2 has facilitated the rise of structure prediction performance to new heights, regularly competitive with experimental structures in the 14th Critical Assessment of Protein Structure Prediction (CASP14). To provide a comprehensive understanding and guide future research in the field of protein structure prediction for researchers, this review describes various methodologies, assessments, and databases in protein structure prediction, including traditionally used protein structure prediction methods, such as template-based modeling (TBM) and template-free modeling (FM) approaches; recently developed deep learning-based methods, such as contact/distance-guided methods, end-to-end folding methods, and protein language model (PLM)-based methods; multi-domain protein structure prediction methods; the CASP experiments and related assessments; and the recently released AlphaFold Protein Structure Database (AlphaFold DB). We discuss their advantages, disadvantages, and application scopes, aiming to provide researchers with insights through which to understand the limitations, contexts, and effective selections of protein structure prediction methods in protein-related fields.

Toward the solution of the protein structure prediction problem

Since Anfinsen demonstrated that the information encoded in a protein's amino acid sequence determines its structure in 1973, solving the protein structure prediction problem has been the Holy Grail of structural biology. The goal of protein structure prediction approaches is to utilize computational modeling to determine the spatial location of every atom in a protein molecule starting from only its amino acid sequence. Depending on whether homologous structures can be found in the Protein Data Bank (PDB), structure prediction methods have been historically categorized as template-based modeling (TBM) or template-free modeling (FM) approaches. Until recently, TBM has been the most reliable approach to predicting protein structures, and in the absence of reliable templates, the modeling accuracy sharply declines. Nevertheless, the results of the most recent community-wide assessment of protein structure prediction experiment (CASP14) have demonstrated that the protein structure prediction problem can be largely solved through the use of end-to-end deep machine learning techniques, where correct folds could be built for nearly all single-domain proteins without using the PDB templates. Critically, the model quality exhibited little correlation with the quality of available template structures, as well as the number of sequence homologs detected for a given target protein. Thus, the implementation of deep-learning techniques has essentially broken through the 50-year-old modeling border between TBM and FM approaches and has made the success of high-resolution structure prediction significantly less dependent on template availability in the PDB library.

Computational methods in drug discovery

The process for drug discovery and development is challenging, time consuming and expensive. Computer-aided drug discovery (CADD) tools can act as a virtual shortcut, assisting in the expedition of this long process and potentially reducing the cost of research and development. Today CADD has become an effective and indispensable tool in therapeutic development. The human genome project has made available a substantial amount of sequence data that can be used in various drug discovery projects. Additionally, increasing knowledge of biological structures, as well as increasing computer power have made it possible to use computational methods effectively in various phases of the drug discovery and development pipeline. The importance of in silico tools is greater than ever before and has advanced pharmaceutical research. Here we present an overview of computational methods used in different facets of drug discovery and highlight some of the recent successes. In this review, both structure-based and ligand-based drug discovery methods are discussed. Advances in virtual high-throughput screening, protein structure prediction methods, protein-ligand docking, pharmacophore modeling and QSAR techniques are reviewed.

Ab Initio structure prediction for Escherichia coli: towards genome-wide protein structure modeling and fold assignment

Genome-wide protein structure prediction and structure-based function annotation have been a long-term goal in molecular biology but not yet become possible due to difficulties in modeling distant-homology targets. We developed a hybrid pipeline combining ab initio folding and template-based modeling for genome-wide structure prediction applied to the Escherichia coli genome. The pipeline was tested on 43 known sequences, where QUARK-based ab initio folding simulation generated models with TM-score 17% higher than that by traditional comparative modeling methods. For 495 unknown hard sequences, 72 are predicted to have a correct fold (TM-score > 0.5) and 321 have a substantial portion of structure correctly modeled (TM-score > 0.35). 317 sequences can be reliably assigned to a SCOP fold family based on structural analogy to existing proteins in PDB. The presented results, as a case study of E. coli, represent promising progress towards genome-wide structure modeling and fold family assignment using state-of-the-art ab initio folding algorithms.

Optimizing structural modeling for a specific protein scaffold: knottins or inhibitor cystine knots

Background

Knottins are small, diverse and stable proteins with important drug design potential. They can be classified in 30 families which cover a wide range of sequences (1621 sequenced), three-dimensional structures (155 solved) and functions (> 10). Inter knottin similarity lies mainly between 15% and 40% sequence identity and 1.5 to 4.5 Å backbone deviations although they all share a tightly knotted disulfide core. This important variability is likely to arise from the highly diverse loops which connect the successive knotted cysteines. The prediction of structural models for all knottin sequences would open new directions for the analysis of interaction sites and to provide a better understanding of the structural and functional organization of proteins sharing this scaffold.

Results

We have designed an automated modeling procedure for predicting the three-dimensionnal structure of knottins. The different steps of the homology modeling pipeline were carefully optimized relatively to a test set of knottins with known structures: template selection and alignment, extraction of structural constraints and model building, model evaluation and refinement. After optimization, the accuracy of predicted models was shown to lie between 1.50 and 1.96 Å from native structures at 50% and 10% maximum sequence identity levels, respectively. These average model deviations represent an improvement varying between 0.74 and 1.17 Å over a basic homology modeling derived from a unique template. A database of 1621 structural models for all known knottin sequences was generated and is freely accessible from our web server at http://knottin.cbs.cnrs.fr. Models can also be interactively constructed from any knottin sequence using the structure prediction module Knoter1D3D available from our protein analysis toolkit PAT at http://pat.cbs.cnrs.fr.

Conclusions

This work explores different directions for a systematic homology modeling of a diverse family of protein sequences. In particular, we have shown that the accuracy of the models constructed at a low level of sequence identity can be improved by 1) a careful optimization of the modeling procedure, 2) the combination of multiple structural templates and 3) the use of conserved structural features as modeling restraints.

The N-terminal domains of neuregulin 1 confer signal attenuation

Degradation of activated ERBB receptors is an important mechanism for signal attenuation. However, compared with epidermal growth factor (EGF) receptor, the ERBB2/ERBB3 signaling pair is considered to be attenuation-deficient. The ERBB2/ERBB3 ligands of the neuregulin family rely on an EGF-like domain for signaling and are generated from larger membrane-bound precursors. In contrast to EGF, which is processed to yield a 6-kDa peptide ligand, mature neuregulins retain a variety of segments N-terminal to the EGF-like domain. Here we evaluate the role of the N-terminal domain of neuregulin 1 in signaling and turnover of ERBB2/ERBB3. Our data suggest that whereas the EGF-like domain of neuregulin 1 is required and sufficient for the formation of active receptor heterodimers, the presence of the N-terminal Ig-like domain is required for efficient signal attenuation. This manifests itself for both ERBB2 and ERBB3 but is more pronounced and coupled directly to degradation for ERBB3. When stimulated with only the EGF-like domain, ERBB3 shows degradation rates comparable with constitutive turnover, but stimulation with full-length neuregulin 1 resulted in receptor degradation at rates that are comparable with activated EGF receptor. Most of the enhancement in down-regulation was maintained after replacing the Ig-like domain with a thioredoxin protein of comparable size but different amino acid composition, suggesting that the physical presence but not specific properties of the Ig-like domain are needed. This sequence-independent effect of the N-terminal domain correlates with an enhanced ability of full-size neuregulin 1 to disrupt higher order oligomers of the ERBB3 extracellular domains in vitro.

The structure of an ancient conserved domain establishes a structural basis for stable histidine phosphorylation and identifies a new family of adenosine-specific kinases

Phosphorylation of both small molecules and proteins plays a central role in many biological processes. In proteins, phosphorylation most commonly targets the oxygen atoms of Ser, Thr, and Tyr. In contrast, stably phosphorylated His residues are rarely found, due to the lability of the N-P bond, and histidine phosphorylation features most often in transient processes. Here we present the crystal structure of a protein of previously unknown function, which proves to contain a stably phosphorylated histidine residue. The protein is the product of open reading frame PAE2307, from the hyperthermophilic archaeon Pyrobaculum aerophilum, and is representative of a highly conserved protein family found in archaea and bacteria. The crystal structure of PAE2307, solved at 1.45-A resolution (R = 0.208, R(free) = 0.227), forms a remarkably tightly associated hexamer. The phosphorylated histidine at the proposed active site, pHis85, occupies a cavity that is at the interface between two subunits and contains a number of fully conserved residues. Stable phosphorylation is attributed to favorable hydrogen bonding of the phosphoryl group and a salt bridge with pHis85 that provides electronic stabilization. In silico modeling suggested that the protein may function as an adenosine kinase, a conclusion that is supported by in vitro assays of adenosine binding, using fluorescence spectroscopy, and crystallographic visualization of an adenosine complex of PAE2307 at 2.25-A resolution.

LiveBench-8: the large-scale, continuous assessment of automated protein structure prediction

We present the results of the evaluation of the latest LiveBench-8 experiment. These results provide a snapshot view of the state of the art in automated protein structure prediction, just before the 2004 CAFASP-4/CASP-6 experiments begin. The last CAFASP/CASP experiments demonstrated that automated meta-predictors entail a significant advance in the field, already challenging most human expert predictors. LiveBench-8 corroborates the superior performance of meta-predictors, which are able to produce useful predictions for over one-half of the test targets. More importantly, LiveBench-8 identifies a handful of recently developed autonomous (nonmeta) servers that perform at the very top, suggesting that further progress in the individual methods has recently been obtained.

Microbial genomes have over 72% structure assignment by the threading algorithm PROSPECTOR_Q

The genome scale threading of five complete microbial genomes is revisited using our state-of-the-art threading algorithm, PROSPECTOR_Q. Considering that structure assignment to an ORF could be useful for predicting biochemical function as well as for analyzing pathways, it is important to assess the current status of genome scale threading. The fraction of ORFs to which we could assign protein structures with a reasonably good confidence level to each genome sequences is over 72%, which is significantly higher than earlier studies. Using the assigned structures, we have predicted the function of several ORFs through "single-function" template structures, obtained from an analysis of the relationship between protein fold and function. The fold distribution of the genomes and the effect of the number of homologous sequences on structure assignment are also discussed.

Automated structure prediction of weakly homologous proteins on a genomic scale

We have developed TASSER, a hierarchical approach to protein structure prediction that consists of template identification by threading, followed by tertiary structure assembly via the rearrangement of continuous template fragments guided by an optimized C(alpha) and side-chain-based potential driven by threading-based, predicted tertiary restraints. TASSER was applied to a comprehensive benchmark set of 1,489 medium-sized proteins in the Protein Data Bank. With homologues excluded, in 927 cases, the templates identified by our threading algorithm PROSPECTOR_3 have a rms deviation from native <6.5 A with approximately 80% alignment coverage. After template reassembly, this number increases to 1,172. This shows significant and systematic improvement of the final models with respect to the initial template alignments. Furthermore, significant improvements in loop modeling are demonstrated. We then apply TASSER to the 1,360 medium-sized ORFs in the Escherichia coli genome; approximately 920 can be predicted with high accuracy based on confidence criteria established in the Protein Data Bank benchmark. These results from our unprecedented comprehensive folding benchmark on all protein categories provide a reliable basis for the application of TASSER to structural genomics, especially to proteins of low sequence identity to solved protein structures.

Automated structure prediction of weakly homologous proteins on a genomic scale

We have developed TASSER, a hierarchical approach to protein structure prediction that consists of template identification by threading, followed by tertiary structure assembly via the rearrangement of continuous template fragments guided by an optimized C(alpha) and side-chain-based potential driven by threading-based, predicted tertiary restraints. TASSER was applied to a comprehensive benchmark set of 1,489 medium-sized proteins in the Protein Data Bank. With homologues excluded, in 927 cases, the templates identified by our threading algorithm PROSPECTOR_3 have a rms deviation from native <6.5 A with approximately 80% alignment coverage. After template reassembly, this number increases to 1,172. This shows significant and systematic improvement of the final models with respect to the initial template alignments. Furthermore, significant improvements in loop modeling are demonstrated. We then apply TASSER to the 1,360 medium-sized ORFs in the Escherichia coli genome; approximately 920 can be predicted with high accuracy based on confidence criteria established in the Protein Data Bank benchmark. These results from our unprecedented comprehensive folding benchmark on all protein categories provide a reliable basis for the application of TASSER to structural genomics, especially to proteins of low sequence identity to solved protein structures.

Structural characterization of genomes by large scale sequence-structure threading

Background

Using sequence-structure threading we have conducted structural characterization of complete proteomes of 37 archaeal, bacterial and eukaryotic organisms (including worm, fly, mouse and human) totaling 167,888 genes.

Results

The reported data represent first rather general evaluation of performance of full sequence-structure threading on multiple genomes providing opportunity to evaluate its general applicability for large scale studies.

According to the estimated results the sequence-structure threading has assigned protein folds to more then 60% of eukaryotic, 68% of archaeal and 70% of bacterial proteomes.

The repertoires of protein classes, architectures, topologies and homologous superfamilies (according to the CATH 2.4 classification) have been established for distant organisms and superkingdoms. It has been found that the average abundance of CATH classes decreases from "alpha and beta" to "mainly beta", followed by "mainly alpha" and "few secondary structures".

3-Layer (aba) Sandwich has been characterized as the most abundant protein architecture and Rossman fold as the most common topology.

Conclusion

The analysis of genomic occurrences of CATH 2.4 protein homologous superfamilies and topologies has revealed the power-law character of their distributions. The corresponding double logarithmic "frequency – genomic occurrence" dependences characteristic of scale-free systems have been established for individual organisms and for three superkingdoms.

Supplementary materials to this works are available at [1].

3DS3 and 3DS5 3D-SHOTGUN meta-predictors in CAFASP3

The performance of the 3DS3 and 3DS5 3D-SHOTGUN meta-predictors in CAFASP3 is reported. The 3D-SHOTGUN meta-predictors are fully automatic fold recognition servers that attempt to incorporate into the prediction process a number of successful strategies that human predictors often apply. Namely, the input to 3D-SHOTGUN are the top five models predicted by a number of independent fold recognition servers and its output are hybrid models, assembled by using the recurrent structural information from the input models. The resulting hybrid models are, on average, more accurate and more complete than the input models. When evaluated on a large set of prediction targets, the 3D-SHOTGUN servers show increased sensitivities and significantly better specificities. For CAFASP3, the 3DS3 and 3DS3 and 3DS5 used a preliminary implementation of the 3D-SHOTGUN method, which lacked a refinement step. Although this did not have a significant effect on the easier targets, for the hardest prediction targets, where the input models had significant structural conflicts, the 3D-SHOTGUN models contained a number of non-native-like features such as fragmentation and overlaps. The CAFASP3 evaluation identified the 3D-SHOTGUN meta-predictors within the top three most sensitive and most specific servers. A fully automated refinement step to the 3D-SHOTGUN method is currently being implemented, and preliminary results indicate that in addition to "cleaning up" such undesirable features, it is able to further increase the accuracy of the resulting models.

Refinement of homology-based protein structures by molecular dynamics simulation techniques

The use of classical molecular dynamics simulations, performed in explicit water, for the refinement of structural models of proteins generated ab initio or based on homology has been investigated. The study involved a test set of 15 proteins that were previously used by Baker and coworkers to assess the efficiency of the ROSETTA method for ab initio protein structure prediction. For each protein, four models generated using the ROSETTA procedure were simulated for periods of between 5 and 400 nsec in explicit solvent, under identical conditions. In addition, the experimentally determined structure and the experimentally derived structure in which the side chains of all residues had been deleted and then regenerated using the WHATIF program were simulated and used as controls. A significant improvement in the deviation of the model structures from the experimentally determined structures was observed in several cases. In addition, it was found that in certain cases in which the experimental structure deviated rapidly from the initial structure in the simulations, indicating internal strain, the structures were more stable after regenerating the side-chain positions. Overall, the results indicate that molecular dynamics simulations on a tens to hundreds of nanoseconds time scale are useful for the refinement of homology or ab initio models of small to medium-size proteins.

The 2002 Olympic Games of protein structure prediction

The summer of every even year is considered by the protein structure prediction community as the Olympic Games season, because in addition to a number of continuous benchmarking experiments such as LiveBench, much effort is invested in the blind prediction experiments CASP and CAFASP. Here we report the major advances registered in the field since the last Games of 2000, as measured by the recently completed LiveBench-4 experiment. These results provide a timely measure of the capabilities of current methods and of their expected performance in the upcoming CASP-5 and CAFASP-3 experiments. We also describe the initiation of the two new, community-wide experiments, PDB-CAFASP and MR-CAFASP. These new experiments extend the scope of previous efforts and may have important implications for structural genomics.

Global perspectives on proteins: comparing genomes in terms of folds, pathways and beyond

The sequencing of complete genomes provides us with a global view of all the proteins in an organism. Proteomic analysis can be done on a purely sequence-based level, with a focus on finding homologues and grouping them into families and clusters of orthologs. However, incorporating protein structure into this analysis provides valuable simplification; it allows one to collect together very distantly related sequences, thus condensing the proteome into a minimal number of 'parts.' We describe issues related to surveying proteomes in terms of structural parts, including methods for fold assignment and formats for comparisons (eg top-10 lists and whole-genome trees), and show how biases in the databases and in sampling can affect these surveys. We illustrate our main points through a case study on the unique protein properties evident in many thermophile genomes (eg more salt bridges). Finally, we discuss metabolic pathways as an even greater simplification of genomes. In comparison to folds these allow the organization of many more genes into coherent systems, yet can nevertheless be understood in many of the same terms.

CAFASP2: the second critical assessment of fully automated structure prediction methods

The results of the second Critical Assessment of Fully Automated Structure Prediction (CAFASP2) are presented. The goals of CAFASP are to (i) assess the performance of fully automatic web servers for structure prediction, by using the same blind prediction targets as those used at CASP4, (ii) inform the community of users about the capabilities of the servers, (iii) allow human groups participating in CASP to use and analyze the results of the servers while preparing their nonautomated predictions for CASP, and (iv) compare the performance of the automated servers to that of the human-expert groups of CASP. More than 30 servers from around the world participated in CAFASP2, covering all categories of structure prediction. The category with the largest participation was fold recognition, where 24 CAFASP servers filed predictions along with 103 other CASP human groups. The CAFASP evaluation indicated that it is difficult to establish an exact ranking of the servers because the number of prediction targets was relatively small and the differences among many servers were also small. However, roughly a group of five "best" fold recognition servers could be identified. The CASP evaluation identified the same group of top servers albeit with a slightly different relative order. Both evaluations ranked a semiautomated method named CAFASP-CONSENSUS, that filed predictions using the CAFASP results of the servers, above any of the individual servers. Although the predictions of the CAFASP servers were available to human CASP predictors before the CASP submission deadline, the CASP assessment identified only 11 human groups that performed better than the best server. Furthermore, about one fourth of the top 30 performing groups corresponded to automated servers. At least half of the top 11 groups corresponded to human groups that also had a server in CAFASP or to human groups that used the CAFASP results to prepare their predictions. In particular, the CAFASP-CONSENSUS group was ranked 7. This shows that the automated predictions of the servers can be very helpful to human predictors. We conclude that as servers continue to improve, they will become increasingly important in any prediction process, especially when dealing with genome-scale prediction tasks. We expect that in the near future, the performance difference between humans and machines will continue to narrow and that fully automated structure prediction will become an effective companion and complement to experimental structural genomics.

Structural genomics analysis: characteristics of atypical, common, and horizontally transferred folds

We conducted a structural genomics analysis of the folds and structural superfamilies in the first 20 completely sequenced genomes by focusing on the patterns of fold usage and trying to identify structural characteristics of typical and atypical folds. We assigned folds to sequences using PSI-blast, run with a systematic protocol to reduce the amount of computational overhead. On average, folds could be assigned to about a fourth of the ORFs in the genomes and about a fifth of the amino acids in the proteomes. More than 80% of all the folds in the SCOP structural classification were identified in one of the 20 organisms, with worm and E. coli having the largest number of distinct folds. Folds are particularly effective at comprehensively measuring levels of gene duplication, because they group together even very remote homologues. Using folds, we find the average level of duplication varies depending on the complexity of the organism, ranging from 2.4 in M. genitalium to 32 for the worm, values significantly higher than those observed based purely on sequence similarity. We rank the common folds in the 20 organisms, finding that the top three are the P-loop NTP hydrolase, the ferrodoxin fold, and the TIM-barrel, and discuss in detail the many factors that affect and bias these rankings. We also identify atypical folds that are "unique" to one of the organisms in our study and compare the characteristics of these folds with the most common ones. We find that common folds tend be more multifunctional and associated with more regular, "symmetrical" structures than the unique ones. In addition, many of the unique folds are associated with proteins involved in cell defense (e.g., toxins). We analyze specific patterns of fold occurrence in the genomes by associating some of them with instances of horizontal transfer and others with gene loss. In particular, we find three possible examples of transfer between archaea and bacteria and six between eukarya and bacteria. We make available our detailed results at http://genecensus.org/20.

Ab initio protein structure prediction on a genomic scale: application to the Mycoplasma genitalium genome

An ab initio protein structure prediction procedure, TOUCHSTONE, was applied to all 85 small proteins of the Mycoplasma genitalium genome. TOUCHSTONE is based on a Monte Carlo refinement of a lattice model of proteins, which uses threading-based tertiary restraints. Such restraints are derived by extracting consensus contacts and local secondary structure from at least weakly scoring structures that, in some cases, can lack any global similarity to the sequence of interest. Selection of the native fold was done by using the convergence of the simulation from two different conformational search schemes and the lowest energy structure by a knowledge-based atomic-detailed potential. Among the 85 proteins, for 34 proteins with significant threading hits, the template structures were reasonably well reproduced. Of the remaining 51 proteins, 29 proteins converged to five or fewer clusters. In the test set, 84.8% of the proteins that converged to five or fewer clusters had a correct fold among the clusters. If this statistic is simply applied, 24 proteins (84.8% of the 29 proteins) may have correct folds. Thus, the topology of a total of 58 proteins probably has been correctly predicted. Based on these results, ab initio protein structure prediction is becoming a practical approach.

Use of covariance analysis for the prediction of structural domain boundaries from multiple protein sequence alignments

Current methods for identification of domains within protein sequences require either structural information or the identification of homologous domain sequences in different sequence contexts. Knowledge of structural domain boundaries is important for fold recognition experiments and structural determination by X-ray crystallography or nuclear magnetic resonance spectroscopy using the divide-and-conquer approach. Here, a new and conceptually simple method for the identification of structural domain boundaries in multiple protein sequence alignments is presented. Analysis of covariance at positions within the alignment is first used to predict 3D contacts. By the nature of the domain as an independent folding unit, inter-domain predicted contacts are fewer than intra-domain predicted contacts. By analysing all possible domain boundaries and constructing a smoothed profile of predicted contact density (PCD), true structural domain boundaries are predicted as local profile minima associated with low PCD. A training data set is constructed from 52 non-homologous two-domain protein sequences of known 3D structure and used to determine optimal parameters for the profile analysis. The alignments in the training data set contained 48 +/- 17 (mean +/- SD) sequences and lengths of 257 +/- 121 residues. Of the 47 alignments yielding predictions, 35% of true domain boundaries are predicted to within 15 amino acids by the local profile minimum with the lowest profile value. Including predictions from the second- and third-lowest local minima increases the correct domain boundary coverage to 60%, whereas the lowest five local minima cover 79% of correct domain boundaries. Through further profile analysis, criteria are presented which reliably identify subsets of more accurate predictions. Retrospective analysis of CASP3 targets shows predictions of sufficient accuracy to enable dramatically improved fold recognition results. Finally, a prediction is made for geminivirus AL1 protein which is in full agreement with biochemical data, yielding a plausible, novel threading result.

MODBASE, a database of annotated comparative protein structure models

MODBASE (http://guitar.rockefeller.edu/modbase) is a relational database of annotated comparative protein structure models for all available protein sequences matched to at least one known protein structure. The models are calculated by MODPIPE, an automated modeling pipeline that relies on PSI-BLAST, IMPALA and MODELLER. MODBASE uses the MySQL relational database management system for flexible and efficient querying, and the MODVIEW Netscape plugin for viewing and manipulating multiple sequences and structures. It is updated regularly to reflect the growth of the protein sequence and structure databases, as well as improvements in the software for calculating the models. For ease of access, MODBASE is organized into different datasets. The largest dataset contains models for domains in 304 517 out of 539 171 unique protein sequences in the complete TrEMBL database (23 March 2001); only models based on significant alignments (PSI-BLAST E-value < 10(-4)) and models assessed to have the correct fold are included. Other datasets include models for target selection and structure-based annotation by the New York Structural Genomics Research Consortium, models for prediction of genes in the Drosophila melanogaster genome, models for structure determination of several ribosomal particles and models calculated by the MODWEB comparative modeling web server.

Protein structure prediction and structural genomics

Genome sequencing projects are producing linear amino acid sequences, but full understanding of the biological role of these proteins will require knowledge of their structure and function. Although experimental structure determination methods are providing high-resolution structure information about a subset of the proteins, computational structure prediction methods will provide valuable information for the large fraction of sequences whose structures will not be determined experimentally. The first class of protein structure prediction methods, including threading and comparative modeling, rely on detectable similarity spanning most of the modeled sequence and at least one known structure. The second class of methods, de novo or ab initio methods, predict the structure from sequence alone, without relying on similarity at the fold level between the modeled sequence and any of the known structures. In this Viewpoint, we begin by describing the essential features of the methods, the accuracy of the models, and their application to the prediction and understanding of protein function, both for single proteins and on the scale of whole genomes. We then discuss the important role that protein structure prediction methods play in the growing worldwide effort in structural genomics.

Amino acids 257 to 288 of mouse p48 control the cooperation of polyomavirus large T antigen, replication protein A, and DNA polymerase alpha-primase to synthesize DNA in vitro

Although p48 is the most conserved subunit of mammalian DNA polymerase alpha-primase (pol-prim), the polypeptide is the major species-specific factor for mouse polyomavirus (PyV) DNA replication. Human and murine p48 contain two regions (A and B) that show significantly lower homology than the rest of the protein. Chimerical human-murine p48 was prepared and coexpressed with three wild-type subunits of pol-prim, and four subunit protein complexes were purified. All enzyme complexes synthesized DNA on single-stranded (ss) DNA and replicated simian virus 40 DNA. Although the recombinant protein complexes physically interacted with PyV T antigen (Tag), we determined that the murine region A mediates the species specificity of PyV DNA replication in vitro. More precisely, the nonconserved phenylalanine 262 of mouse p48 is crucial for this activity, and pol-prim with mutant p48, h-S262F, supports PyV DNA replication in vitro. DNA synthesis on RPA-bound ssDNA revealed that amino acid (aa) 262, aa 266, and aa 273 to 288 are involved in the functional cooperation of RPA, pol-prim, and PyV Tag.

Review: what can structural classifications reveal about protein evolution?

In this article we present a review of the methods used for comparing and classifying protein structures. We discuss the hierarchies and populations of fold groups and evolutionary families in some of the major classifications and we consider some of the problems confronting any general analyses of structural evolution in protein families. We also review some more recent analyses that have expanded these classifications by identifying sequence relatives in the genomes and thereby reveal interesting trends in fold usage and recurrence.

Fold predictions for bacterial genomes

Fold assignments for newly sequenced genomes belong to the most important and interesting applications of the booming field of protein structure prediction. We present a brief survey and a discussion of such assignments completed to date, using as an example several fold assignment projects for proteins from the Escherichia coli genome. This review focuses on steps that are necessary to go beyond the simple assignment projects and into the development of tools extending our understanding of functions of proteins in newly sequenced genomes. This paper also discusses several problems seldom addressed in the literature, such as the problem of domain prediction and complementary predictions (e.g., transmembrane regions and flexible regions) and cross-correlation of predictions from different servers. The influence of sequence and structure database growth on prediction success is also addressed. Finally, we discuss the perspectives of the field in the context of massive sequence and structure determination projects, as well as the development of novel prediction methods.

An integrated approach to structural genomics

Structural genomics aims at determining a set of protein structures that will represent all domain folds present in the biosphere. These structures can be used as the basis for the homology modelling of the majority of all remaining protein domains or, indeed, proteins. Structural genomics therefore promises to provide a comprehensive structural description of the protein universe. To achieve this, a broad scientific effort is required. The Berlin-based "Protein Structure Factory" (PSF) plans to contribute to this effort by setting up a local infrastructure for the low-cost, high-throughput analysis of soluble human proteins. In close collaboration with the German Human Genome Project (DHGP) protein-coding genes will be expressed in Escherichia coli or yeast. Affinity-tagged proteins will be purified semi-automatically for biophysical characterization and structure analysis by X-ray diffraction methods and NMR spectroscopy. In all steps of the structure analysis process, possibilities for automation, parallelization and standardization will be explored. Major new facilities that are created for the PSF include a robotic station for large-scale protein crystallization, an NMR center and an experimental station for protein crystallography at the synchrotron storage ring BESSY II in Berlin.

Comparative protein structure modeling of genes and genomes

Comparative modeling predicts the three-dimensional structure of a given protein sequence (target) based primarily on its alignment to one or more proteins of known structure (templates). The prediction process consists of fold assignment, target-template alignment, model building, and model evaluation. The number of protein sequences that can be modeled and the accuracy of the predictions are increasing steadily because of the growth in the number of known protein structures and because of the improvements in the modeling software. Further advances are necessary in recognizing weak sequence-structure similarities, aligning sequences with structures, modeling of rigid body shifts, distortions, loops and side chains, as well as detecting errors in a model. Despite these problems, it is currently possible to model with useful accuracy significant parts of approximately one third of all known protein sequences. The use of individual comparative models in biology is already rewarding and increasingly widespread. A major new challenge for comparative modeling is the integration of it with the torrents of data from genome sequencing projects as well as from functional and structural genomics. In particular, there is a need to develop an automated, rapid, robust, sensitive, and accurate comparative modeling pipeline applicable to whole genomes. Such large-scale modeling is likely to encourage new kinds of applications for the many resulting models, based on their large number and completeness at the level of the family, organism, or functional network.

LiveBench-1: continuous benchmarking of protein structure prediction servers

We present a novel, continuous approach aimed at the large-scale assessment of the performance of available fold-recognition servers. Six popular servers were investigated: PDB-Blast, FFAS, T98-lib, GenTHREADER, 3D-PSSM, and INBGU. The assessment was conducted using as prediction targets a large number of selected protein structures released from October 1999 to April 2000. A target was selected if its sequence showed no significant similarity to any of the proteins previously available in the structural database. Overall, the servers were able to produce structurally similar models for one-half of the targets, but significantly accurate sequence-structure alignments were produced for only one-third of the targets. We further classified the targets into two sets: easy and hard. We found that all servers were able to find the correct answer for the vast majority of the easy targets if a structurally similar fold was present in the server's fold libraries. However, among the hard targets--where standard methods such as PSI-BLAST fail--the most sensitive fold-recognition servers were able to produce similar models for only 40% of the cases, half of which had a significantly accurate sequence-structure alignment. Among the hard targets, the presence of updated libraries appeared to be less critical for the ranking. An "ideally combined consensus" prediction, where the results of all servers are considered, would increase the percentage of correct assignments by 50%. Each server had a number of cases with a correct assignment, where the assignments of all the other servers were wrong. This emphasizes the benefits of considering more than one server in difficult prediction tasks. The LiveBench program (http://BioInfo.PL/LiveBench) is being continued, and all interested developers are cordially invited to join.

Proteomics of Mycoplasma genitalium: identification and characterization of unannotated and atypical proteins in a small model genome

We present the results of a comprehensive analysis of the proteome of Mycoplasma genitalium (MG), the smallest autonomously replicating organism that has been completely sequenced. Our aim was to identify and characterize all soluble proteins in MG that are structurally and functionally uncharacterized. We were particularly interested in identifying proteins that differed significantly from typical globular proteins, for example, proteins which are unstructured in the absence of a 'partner' molecule or those that exhibit unusual thermodynamic properties. This work is complementary to other structural genomics projects whose primary aim is to determine the three-dimensional structures of proteins with unknown folds. We have identified all the full-length open reading frames (ORFs) in MG that have no homologs of known structure and are of unknown function. Twenty-five of the total 483 ORFs fall into this category and we have expressed, purified and characterized 11 of them. We have used circular dichroism (CD) to rapidly investigate their biophysical properties. Our studies reveal that these proteins have a wide range of structures varying from highly helical to partially structured to unfolded or random coil. They also display a variety of thermodynamic properties ranging from cooperative unfolding to no detectable unfolding upon thermal denaturation. Several of these proteins are highly conserved from mycoplasma to man. Further information about target selection and CD results is available at http://bioinfo.mbb.yale.edu/genome

Large-scale comparison of protein sequence alignment algorithms with structure alignments

Sequence alignment programs such as BLAST and PSI-BLAST are used routinely in pairwise, profile-based, or intermediate-sequence-search (ISS) methods to detect remote homologies for the purposes of fold assignment and comparative modeling. Yet, the sequence alignment quality of these methods at low sequence identity is not known. We have used the CE structure alignment program (Shindyalov and Bourne, Prot Eng 1998;11:739) to derive sequence alignments for all superfamily and family-level related proteins in the SCOP domain database. CE aligns structures and their sequences based on distances within each protein, rather than on interprotein distances. We compared BLAST, PSI-BLAST, CLUSTALW, and ISS alignments with the CE structural alignments. We found that global alignments with CLUSTALW were very poor at low sequence identity (<25%), as judged by the CE alignments. We used PSI-BLAST to search the nonredundant sequence database (nr) with every sequence in SCOP using up to four iterations. The resulting matrix was used to search a database of SCOP sequences. PSI-BLAST is only slightly better than BLAST in alignment accuracy on a per-residue basis, but PSI-BLAST matrix alignments are much longer than BLAST's, and so align correctly a larger fraction of the total number of aligned residues in the structure alignments. Any two SCOP sequences in the same superfamily that shared a hit or hits in the nr PSI-BLAST searches were identified as linked by the shared intermediate sequence. We examined the quality of the longest SCOP-query/ SCOP-hit alignment via an intermediate sequence, and found that ISS produced longer alignments than PSI-BLAST searches alone, of nearly comparable per-residue quality. At 10-15% sequence identity, BLAST correctly aligns 28%, PSI-BLAST 40%, and ISS 46% of residues according to the structure alignments. We also compared CE structure alignments with FSSP structure alignments generated by the DALI program. In contrast to the sequence methods, CE and structure alignments from the FSSP database identically align 75% of residue pairs at the 10-15% level of sequence identity, indicating that there is substantial room for improvement in these sequence alignment methods. BLAST produced alignments for 8% of the 10,665 nonimmunoglobulin SCOP superfamily sequence pairs (nearly all <25% sequence identity), PSI-BLAST matched 17% and the double-PSI-BLAST ISS method aligned 38% with E-values <10.0. The results indicate that intermediate sequences may be useful not only in fold assignment but also in achieving more complete sequence alignments for comparative modeling.

Enhanced genome annotation using structural profiles in the program 3D-PSSM

A method (three-dimensional position-specific scoring matrix, 3D-PSSM) to recognise remote protein sequence homologues is described. The method combines the power of multiple sequence profiles with knowledge of protein structure to provide enhanced recognition and thus functional assignment of newly sequenced genomes. The method uses structural alignments of homologous proteins of similar three-dimensional structure in the structural classification of proteins (SCOP) database to obtain a structural equivalence of residues. These equivalences are used to extend multiply aligned sequences obtained by standard sequence searches. The resulting large superfamily-based multiple alignment is converted into a PSSM. Combined with secondary structure matching and solvation potentials, 3D-PSSM can recognise structural and functional relationships beyond state-of-the-art sequence methods. In a cross-validated benchmark on 136 homologous relationships unambiguously undetectable by position-specific iterated basic local alignment search tool (PSI-Blast), 3D-PSSM can confidently assign 18 %. The method was applied to the remaining unassigned regions of the Mycoplasma genitalium genome and an additional 13 regions were assigned with 95 % confidence. 3D-PSSM is available to the community as a web server: http://www.bmm.icnet.uk/servers/3dpssm

Protein structure prediction in the postgenomic era

As the number of completely sequenced genomes rapidly increases, the postgenomic problem of gene function identification becomes ever more pressing. Predicting the structures of proteins encoded by genes of interest is one possible means to glean subtle clues as to the functions of these proteins. There are limitations to this approach to gene identification and a survey of the expected reliability of different protein structure prediction techniques has been undertaken.

Structural genomics of microbes: an objective

A comparison of the genome sequences of more than 20 microorganisms reveals that a large fraction of the genes have unknown functions. Determining the structures of the proteins coded by these genes may provide additional key information in an effort to uncover the molecular functions of such proteins and new protein fold patterns. Using existing technology, it is possible to obtain a complete sequence complement and a near complete structural complement for a small microbial genome. Such information may provide a comprehensive view of a small organism, which, in turn, can serve as a platform for understanding more complex organisms.

Gene content phylogeny of herpesviruses

Clusters of orthologous groups [COGs; Tatusov, R. L., Koonin, E. V. & Lipman, D. J. (1997) Science 278, 631-637] were identified for a set of 13 completely sequenced herpesviruses. Each COG represented a family of gene products conserved across several herpes genomes. These families were defined without using an arbitrary threshold criterion based on sequence similarity. The COG technique was modified so that variable stringency in COG construction was possible. High stringencies identify a core set of highly conserved genes. Varying COG stringency reveals differences in the degree of conservation between functional classes of genes. The COG data were used to construct whole-genome phylogenetic trees based on gene content. These trees agree well with trees based on other methods and are robust when tested by bootstrap analysis. The COG data also were used to construct a reciprocal tree that clustered genes with similar phylogenetic profiles. This clustering may give clues to genes with related functions or with related histories of acquisition and loss during herpesvirus evolution.

From fold to function predictions: an apoptosis regulator protein BID

With the rapidly increasing pace of genome sequencing projects and the resulting flood of predicted amino acid sequences of uncharacterized proteins, protein sequence analysis, and in particular, protein structure prediction is quickly gaining in importance. Prediction algorithms can be used for preliminary annotation of newly sequenced proteins and, at least in some cases, provide insights into their function and specific mode of action. Such annotations for several microbial genomes were performed by several groups and placed in public domain for evaluation. An example presented in this work comes from a related project of structural and functional predictions for proteins involved in the process of controlled cell death (apoptosis). The BID protein belongs to an important class of regulators of apoptosis identified by short sequence motifs. Here, several fold prediction methods are used to build a series of three-dimensional models. Structure analysis of the models with reference to the biological data available allows selection of the most appropriate model. It is found that the most likely structural model of BID is built on the structure of Bcl-X(L). The model is discussed in terms of experimental data on specific proteolytic cleavage of BID and its effect on BID interactions with other proteins and membranes.

From fold recognition to homology modeling: an analysis of protein modeling challenges at different levels of prediction complexity

An analysis of different approaches to protein structure prediction is presented based solely on the range of models submitted to the third Critical Assessment of Protein Structure Prediction (CASP3) conference. CASP conferences evaluate the current state of the art of protein structure prediction by comparing blind prediction efforts of many groups for the same set of target sequences. Target sequences may be highly similar to those with known structure or can be totally (at least superficially) sequentially dissimilar. Techniques applied to those blind predictions (over 40 targets) ranges from a detailed homology prediction to the detection of remote homologues well below a twilight zone of protein sequence similarity. For the CASP3 conference, we have submitted predictions, totaling 35, with various levels of difficulty and complexity. For ten submitted homology targets, eight of them were determined by experiment so far. The RMSD of C-alpha atoms are 1.2-1.7, 2.3, and 4.6-17.9 A for the three easy targets, two hard targets, and three very hard homology targets, respectively. Out of 18-fold recognition predictions available for analysis, we got six correct predictions, five near misses, three tough near misses and four far misses. Here we analyze successes and failures of those predictions in an attempt to identify common problems and common achievements.

Selecting protein targets for structural genomics of Pyrobaculum aerophilum: validating automated fold assignment methods by using binary hypothesis testing

Three-dimensional protein folds were assigned to all ORFs of the recently sequenced genome of the hyperthermophilic archaeon Pyrobaculum aerophilum. Binary hypothesis testing was used to estimate a confidence level for each assignment. A separate test was conducted to assign a probability for whether each sequence has a novel fold-i.e., one that is not yet represented in the experimental database of known structures. Of the 2,130 predicted nontransmembrane proteins in this organism, 916 matched a fold at a cumulative 90% confidence level, and 245 could be assigned at a 99% confidence level. Likewise, 286 proteins were predicted to have a previously unobserved fold with a 90% confidence level, and 14 at a 99% confidence level. These statistically based tools are combined with homology searches against the Online Mendelian Inheritance in Man (OMIM) human genetics database and other protein databases for the selection of attractive targets for crystallographic or NMR structure determination. Results of these studies have been collated and placed at http://www.doe-mbi.ucla.edu/people/parag/P A_HOME/, the University of California, Los Angeles-Department of Energy Pyrobaculum aerophilum web site.

Structural genomics and its importance for gene function analysis

Structural genomics projects aim to solve the experimental structures of all possible protein folds. Such projects entail a conceptual shift from traditional structural biology in which structural information is obtained on known proteins to one in which the structure of a protein is determined first and the function assigned only later. Whereas the goal of converting protein structure into function can be accomplished by traditional sequence motif-based approaches, recent studies have shown that assignment of a protein's biochemical function can also be achieved by scanning its structure for a match to the geometry and chemical identity of a known active site. Importantly, this approach can use low-resolution structures provided by contemporary structure prediction methods. When applied to genomes, structural information (either experimental or predicted) is likely to play an important role in high-throughput function assignment.

MODBASE, a database of annotated comparative protein structure models

MODBASE is a queryable database of annotated comparative protein structure models. The models are derived by MODPIPE, an automated modeling pipeline relying on the programs PSI-BLAST and MODELLER. The database currently contains 3D models for substantial portions of approximately 17 000 proteins from 10 complete genomes, including those of Caenorhabditis elegans, Saccharomyces cerevisiae and Escherichia coli, as well as all the available sequences from Arabidopsis thaliana and Homo sapiens. The database also includes fold assignments and alignments on which the models were based. In addition, special care is taken to assess the quality of the models. ModBase is accessible through a web interface at http://guitar.rockefeller.edu/modbase/

New approaches to rational drug design

Rational drug design has emerged as a powerful technique. In this review, three new developments in rational drug design are explored. These developments include new methods to find binding sites for small molecules on the surface of a protein, the suggestion that the protein environment may change the shape of a protein sufficiently to alter drug design, and the use of data emerging from structural genomics in drug design. Although these are three new and distinct areas, the insights derived from these studies suggest a reason for the observation that similar drugs do not always bind to a protein in the same manner.

Receptor-independent activators of heterotrimeric G-protein signaling pathways

Heterotrimeric G-protein signaling systems are activated via cell surface receptors possessing the seven-membrane span motif. Several observations suggest the existence of other modes of stimulus input to heterotrimeric G-proteins. As part of an overall effort to identify such proteins we developed a functional screen based upon the pheromone response pathway in Saccharomyces cerevisiae. We identified two mammalian proteins, AGS2 and AGS3 (activators of G-protein signaling), that activated the pheromone response pathway at the level of heterotrimeric G-proteins in the absence of a typical receptor. beta-galactosidase reporter assays in yeast strains expressing different Galpha subunits (Gpa1, G(s)alpha, G(i)alpha(2(Gpa1(1-41))), G(i)alpha(3(Gpa1(1-41))), Galpha(16(Gpa1(1-41)))) indicated that AGS proteins selectively activated G-protein heterotrimers. AGS3 was only active in the G(i)alpha(2) and G(i)alpha(3) genetic backgrounds, whereas AGS2 was active in each of the genetic backgrounds except Gpa1. In protein interaction studies, AGS2 selectively associated with Gbetagamma, whereas AGS3 bound Galpha and exhibited a preference for GalphaGDP versus GalphaGTPgammaS. Subsequent studies indicated that the mechanisms of G-protein activation by AGS2 and AGS3 were distinct from that of a typical G-protein-coupled receptor. AGS proteins provide unexpected mechanisms for input to heterotrimeric G-protein signaling pathways. AGS2 and AGS3 may also serve as novel binding partners for Galpha and Gbetagamma that allow the subunits to subserve functions that do not require initial heterotrimer formation.