Automatic generation and evaluation of sparse protein signatures for families of protein structural domains

We identified key residues from the structural alignment of families of protein domains from SCOP which we represented in the form of sparse protein signatures. A signature-generating algorithm (SigGen) was developed and used to automatically identify key residues based on several structural and sequence-based criteria. The capacity of the signatures to detect related sequences from the SWISSPROT database was assessed by receiver operator characteristic (ROC) analysis and jack-knife testing. Test signatures for families from each of the main SCOP classes are described in relation to the quality of the structural alignments, the SigGen parameters used, and their diagnostic performance. We show that automatically generated signatures are potently diagnostic for their family (ROC50 scores typically >0.8), consistently outperform random signatures, and can identify sequence relationships in the "twilight zone" of protein sequence similarity (<40%). Signatures based on 15%-30% of alignment positions occurred most frequently among the best-performing signatures. When alignment quality is poor, sparser signatures perform better, whereas signatures generated from higher-quality alignments of fewer structures require more positions to be diagnostic. Our validation of signatures from the Globin family shows that when sequences from the structural alignment are removed and new signatures generated, the omitted sequences are still detected. The positions highlighted by the signature often correspond (alignment specificity >0.7) to the key positions in the original (non-jack-knifed) alignment. We discuss potential applications of sparse signatures in sequence annotation and homology modeling.     

Template-based recognition of protein fold within the midnight and twilight zones of protein sequence similarity

Most homologous pairs of proteins have no significant sequence similarity to each other and are not identified by direct sequence comparison or profile-based strategies. However, multiple sequence alignments of low similarity homologues typically reveal a limited number of positions that are well conserved despite diversity of function. It may be inferred that conservation at most of these positions is the result of the importance of the contribution of these amino acids to the folding and stability of the protein. As such, these amino acids and their relative positions may define a structural signature. We demonstrate that extraction of this fold template provides the basis for the sequence database to be searched for patterns consistent with the fold, enabling identification of homologs that are not recognized by global sequence analysis. The fold template method was developed to address the need for a tool that could comprehensively search the midnight and twilight zones of protein sequence similarity without reliance on global statistical significance. Manual implementations of the fold template method were performed on three folds--immunoglobulin, c-lectin and TIM barrel. Following proof of concept of the template method, an automated version of the approach was developed. This automated fold template method was used to develop fold templates for 10 of the more populated folds in the SCOP database. The fold template method developed three-dimensional structural motifs or signatures that were able to return a diverse collection of proteins, while maintaining a low false positive rate. Although the results of the manual fold template method were more comprehensive than the automated fold template method, the diversity of the results from the automated fold template method surpassed those of current methods that rely on statistical significance to infer evolutionary relationships among divergent proteins.     

The size distribution of protein families within different types of folds

It is well known that the structure is currently available only for a small fraction of known protein sequences. It is urgent to discover the important features of known protein sequences based on present protein structures. Here, we report a study on the size distribution of protein families within different types of folds. The fold of a protein means the global arrangement of its main secondary structures, both in terms of their relative orientations and their topological connections, which specify a certain biochemical and biophysical aspect. We first search protein families in the structural database SCOP against the sequence-based database Pfam, and acquire a pool of corresponding Pfam families whose structures can be deemed as known. This pool of Pfam families is called the sample space for short. Then the size distributions of protein families involving the sample space, the Pfam database and the SCOP database are obtained. The results indicate that the size distributions of protein families under different kinds of folds abide by similar power-law. Specially, the largest families scatter evenly in different kinds of folds. This may help better understand the relationship of protein sequence, structure and function. We also show that the total of proteins with known structures can be considered a random sample from the whole space of protein sequences, which is an essential but unsettled assumption for related predictions, such as, estimating the number of protein folds in nature. Finally we conclude that about 2957 folds are needed to cover the total Pfam families by a simple method.     

Estimating the prevalence of protein sequences adopting functional enzyme folds

Proteins employ a wide variety of folds to perform their biological functions. How are these folds first acquired? An important step toward answering this is to obtain an estimate of the overall prevalence of sequences adopting functional folds. Since tertiary structure is needed for a typical enzyme active site to form, one way to obtain this estimate is to measure the prevalence of sequences supporting a working active site. Although the immense number of sequence combinations makes wholly random sampling unfeasible, two key simplifications may provide a solution. First, given the importance of hydrophobic interactions to protein folding, it seems likely that the sample space can be restricted to sequences carrying the hydropathic signature of a known fold. Second, because folds are stabilized by the cooperative action of many local interactions distributed throughout the structure, the overall problem of fold stabilization may be viewed reasonably as a collection of coupled local problems. This enables the difficulty of the whole problem to be assessed by assessing the difficulty of several smaller problems. Using these simplifications, the difficulty of specifying a working beta-lactamase domain is assessed here. An alignment of homologous domain sequences is used to deduce the pattern of hydropathic constraints along chains that form the domain fold. Starting with a weakly functional sequence carrying this signature, clusters of ten side-chains within the fold are replaced randomly, within the boundaries of the signature, and tested for function. The prevalence of low-level function in four such experiments indicates that roughly one in 10(64) signature-consistent sequences forms a working domain. Combined with the estimated prevalence of plausible hydropathic patterns (for any fold) and of relevant folds for particular functions, this implies the overall prevalence of sequences performing a specific function by any domain-sized fold may be as low as 1 in 10(77), adding to the body of evidence that functional folds require highly extraordinary sequences.     

AutoSCOP: automated prediction of SCOP classifications using unique pattern-class mappings

MOTIVATION: The sequence patterns contained in the available motif and hidden Markov model (HMM) databases are a valuable source of information for protein sequence annotation. For structure prediction and fold recognition purposes, we computed mappings from such pattern databases to the protein domain hierarchy given by the ASTRAL compendium and applied them to the prediction of SCOP classifications. Our aim is to make highly confident predictions also for non-trivial cases if possible and abstain from a prediction otherwise, and thus to provide a method that can be used as a first step in a pipeline of prediction methods. We describe two successful examples for such pipelines. With the AutoSCOP approach, it is possible to make predictions in a large-scale manner for many domains of the available sequences in the well-known protein sequence databases. RESULTS: AutoSCOP computes unique sequence patterns and pattern combinations for SCOP classifications. For instance, we assign a SCOP superfamily to a pattern found in its members whenever the pattern does not occur in any other SCOP superfamily. Especially on the fold and superfamily level, our method achieves both high sensitivity (above 93%) and high specificity (above 98%) on the difference set between two ASTRAL versions, due to being able to abstain from unreliable predictions. Further, on a harder test set filtered at low sequence identity, the combination with profile-profile alignments improves accuracy and performs comparably even to structure alignment methods. Integrating our method with structure alignment, we are able to achieve an accuracy of 99% on SCOP fold classifications on this set. In an analysis of false assignments of domains from new folds/superfamilies/families to existing SCOP classifications, AutoSCOP correctly abstains for more than 70% of the domains belonging to new folds and superfamilies, and more than 80% of the domains belonging to new families. These findings show that our approach is a useful additional filter for SCOP classification prediction of protein domains in combination with well-known methods such as profile-profile alignment. AVAILABILITY: A web server where users can input their domain sequences is available at http://www.bio.ifi.lmu.de/autoscop.     

A fast SCOP fold classification system using content-based E-Predict algorithm

Background  

Domain experts manually construct the Structural Classification of Protein (SCOP) database to categorize and compare protein structures. Even though using the SCOP database is believed to be more reliable than classification results from other methods, it is labor intensive. To mimic human classification processes, we develop an automatic SCOP fold classification system to assign possible known SCOP folds and recognize novel folds for newly-discovered proteins.     

一个识别蛋白质折叠模式的SVM分类器

蛋白质折叠模式识别是一种分析蛋白质结构的重要方法。以序列相似性较低的蛋白质为训练集,提取蛋白质序列信息频数及疏水性等信息作为折叠类型特征,从SCOP数据库中已分类蛋白质构建1 393种折叠模式的数据集,采用SVM预测蛋白质1 393种折叠模式。封闭测试准确率达99.612 2%,基于SCOP的开放测试准确率达79.632 9%。基于另一个权威测试集的开放测试折叠准确率达64.705 9%,SCOP类准确率达76.470 6%,可以有效地对蛋白质折叠模式进行预测,从而为蛋白质从头预测提供参考。     

The automatic discovery of structural principles describing protein fold space

The study of protein structure has been driven largely by the careful inspection of experimental data by human experts. However, the rapid determination of protein structures from structural-genomics projects will make it increasingly difficult to analyse (and determine the principles responsible for) the distribution of proteins in fold space by inspection alone. Here, we demonstrate a machine-learning strategy that automatically determines the structural principles describing 45 folds. The rules learnt were shown to be both statistically significant and meaningful to protein experts. With the increasing emphasis on high-throughput experimental initiatives, machine-learning and other automated methods of analysis will become increasingly important for many biological problems.     

How a Spatial Arrangement of Secondary Structure Elements Is Dispersed in the Universe of Protein Folds

It has been known that topologically different proteins of the same class sometimes share the same spatial arrangement of secondary structure elements (SSEs). However, the frequency by which topologically different structures share the same spatial arrangement of SSEs is unclear. It is important to estimate this frequency because it provides both a deeper understanding of the geometry of protein folds and a valuable suggestion for predicting protein structures with novel folds. Here we clarified the frequency with which protein folds share the same SSE packing arrangement with other folds, the types of spatial arrangement of SSEs that are frequently observed across different folds, and the diversity of protein folds that share the same spatial arrangement of SSEs with a given fold, using a protein structure alignment program MICAN, which we have been developing. By performing comprehensive structural comparison of SCOP fold representatives, we found that approximately 80% of protein folds share the same spatial arrangement of SSEs with other folds. We also observed that many protein pairs that share the same spatial arrangement of SSEs belong to the different classes, often with an opposing N- to C-terminal direction of the polypeptide chain. The most frequently observed spatial arrangement of SSEs was the 2-layer α/β packing arrangement and it was dispersed among as many as 27% of SCOP fold representatives. These results suggest that the same spatial arrangements of SSEs are adopted by a wide variety of different folds and that the spatial arrangement of SSEs is highly robust against the N- to C-terminal direction of the polypeptide chain.     

Automated identification of protein-ligand interaction features using Inductive Logic Programming: A hexose binding case study

ABSTRACT: BACKGROUND: There is a need for automated methods to learn general features of the interactions of a ligand class with its diverse set of protein receptors. An appropriate machine learning approach is Inductive Logic Programming (ILP), which automatically generates comprehensible rules in addition to prediction. The development of ILP systems whichcan learn rules of the complexity required for studies on protein structure remains a challenge. In this work we use a new ILP system, ProGolem, and demonstrate its performance on learning features of hexose-protein interactions. RESULTS: The rules induced by ProGolem detect interactions mediated by aromatics and by planar-polar residues, in addition to less common features such as the aromatic sandwich. The rules also reveal a previously unreported dependency for residues CYS and LEU. They also specify interactions involving aromatic and hydrogen bonding residues. CONCLUSIONS: In addition to confirming literature results, ProGolem's model has a 10-fold cross-validated predictive accuracy that is superior, at the 95% confidence level, to another ILP system previously used to study protein/hexose interactions and is comparable with state-of-the-art statistical learners.     

Visualization of protein fold space via nonmetric multidimensional scaling

In this paper, a 3D map of protein fold space was produced using Dali structure alignment and nonmetric multidimensional scaling. The fold space comprises four radial clusters, which correspond to the four classes of SCOP. The overall structure of the protein fold space is largely determined by three factors: secondary structure composition, topology of beta sheet, and domain size.     

The relationship between protein structure and function: a comprehensive survey with application to the yeast genome.

For most proteins in the genome databases, function is predicted via sequence comparison. In spite of the popularity of this approach, the extent to which it can be reliably applied is unknown. We address this issue by systematically investigating the relationship between protein function and structure. We focus initially on enzymes functionally classified by the Enzyme Commission (EC) and relate these to by structurally classified domains the SCOP database. We find that the major SCOP fold classes have different propensities to carry out certain broad categories of functions. For instance, alpha/beta folds are disproportionately associated with enzymes, especially transferases and hydrolases, and all-alpha and small folds with non-enzymes, while alpha+beta folds have an equal tendency either way. These observations for the database overall are largely true for specific genomes. We focus, in particular, on yeast, analyzing it with many classifications in addition to SCOP and EC (i.e. COGs, CATH, MIPS), and find clear tendencies for fold-function association, across a broad spectrum of functions. Analysis with the COGs scheme also suggests that the functions of the most ancient proteins are more evenly distributed among different structural classes than those of more modern ones. For the database overall, we identify the most versatile functions, i.e. those that are associated with the most folds, and the most versatile folds, associated with the most functions. The two most versatile enzymatic functions (hydro-lyases and O-glycosyl glucosidases) are associated with seven folds each. The five most versatile folds (TIM-barrel, Rossmann, ferredoxin, alpha-beta hydrolase, and P-loop NTP hydrolase) are all mixed alpha-beta structures. They stand out as generic scaffolds, accommodating from six to as many as 16 functions (for the exceptional TIM-barrel). At the conclusion of our analysis we are able to construct a graph giving the chance that a functional annotation can be reliably transferred at different degrees of sequence and structural similarity. Supplemental information is available from http://bioinfo.mbb.yale.edu/genome/foldfunc++ +.     

Comparison of sequence and structure-based datasets for nonredundant structural data mining

Structural data mining studies attempt to deduce general principles of protein structure from solved structures deposited in the protein data bank (PDB). The entire database is unsuitable for such studies because it is not representative of the ensemble of protein folds. Given that novel folds continue to be unearthed, some folds are currently unrepresented in the PDB while other folds are overrepresented. Overrepresentation can easily be avoided by filtering the dataset. PDB_SELECT is a well-used representative subset of the PDB that has been deduced by sequence comparison. Specifically, structures with sequences that exhibit a pairwise sequence identity above a threshold value are weeded from the dataset. Although length criteria for pairwise alignments have a structural basis, this automated method of pruning is essentially sequence-based and runs into problems in the twilight zone, possibly resulting in some folds being overrepresented. The value-added structure databases SCOP and CATH are also a potential source of a nonredundant dataset. Here we compare the sequence-derived dataset PDB_SELECT with the structural databases SCOP (Structural Classification Of Proteins) and CATH (Class-Architecture-Topology-Homology). We show that some folds remain overrepresented in the PDB_SELECT dataset while other folds are not represented at all. However, SCOP and CATH also have their own problems such as the labor-intensiveness of the update process and the problem of determining whether all folds are equally or sufficiently distant. We discuss areas where further work is required.     

A unifold, mesofold, and superfold model of protein fold use.

As more and more protein structures are determined, there is increasing interest in the question of how many different folds have been used in biology. The history of the rate of discovery of new folds and the distribution of sequence families among known folds provide a means of estimating the underlying distribution of fold use. Previous models exploiting these data have led to rather different conclusions on the total number of folds. We present a new model, based on the notion that the folds used in biology fall naturally into three classes: unifolds, that is, folds found only in a single narrow sequence family; mesofolds, found in an intermediate number of families; and the previously noted superfolds, found in many protein families. We show that this model fits the available data well and has predicted the development of SCOP over the past 2 years. The principle implications of the model are as follows: (1) The vast majority of folds will be found in only a single sequence family; (2) the total number of folds is at least 10,000; and (3) 80% of sequence families have one of about 400 folds, most of which are already known.     

The number of protein folds and their distribution over families in nature

Currently, of the 10(6) known protein sequences, only about 10(4) structures have been solved. Based on homologies and similarities, proteins are grouped into different families in which each has a structural prototype, namely, the fold, and some share the same folds. However, the total number of folds and families, and furthermore, the distribution of folds over families in nature, are still an enigma. Here, we report a study on the distribution of folds over families and the total number of folds in nature, using a maximum probability principle and the moment method of estimation. A quadratic relation between the numbers of families and folds is found for the number of families in an interval from 6000 to 30,000. For example, about 2700 folds for 23,100 families are obtained, among them about 33 superfolds, including more than 100 families each, and the largest superfold comprises about 800 families. Our results suggest that although the majority of folds have only a single family per fold, a considerably larger number of folds include many more families each than in the database, and the distribution of folds over families in nature differs markedly from the sampled distribution. The long tail of fold distribution is first estimated in this article. The results fit the data for different versions of the structural classification of proteins (SCOP) excellently, and the goodness-of-fit tests strongly support the results. In addition, the method of directly "enlarging" the sample to the population may be useful in inferring distributions of species in different fields.     

Multi-class protein fold recognition using support vector machines and neural networks

MOTIVATION: Protein fold recognition is an important approach to structure discovery without relying on sequence similarity. We study this approach with new multi-class classification methods and examined many issues important for a practical recognition system. RESULTS: Most current discriminative methods for protein fold prediction use the one-against-others method, which has the well-known 'False Positives' problem. We investigated two new methods: the unique one-against-others and the all-against-all methods. Both improve prediction accuracy by 14-110% on a dataset containing 27 SCOP folds. We used the Support Vector Machine (SVM) and the Neural Network (NN) learning methods as base classifiers. SVMs converges fast and leads to high accuracy. When scores of multiple parameter datasets are combined, majority voting reduces noise and increases recognition accuracy. We examined many issues involved with large number of classes, including dependencies of prediction accuracy on the number of folds and on the number of representatives in a fold. Overall, recognition systems achieve 56% fold prediction accuracy on a protein test dataset, where most of the proteins have below 25% sequence identity with the proteins used in training.     

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.     

Decision tree based information integration for automated protein classification

We propose a novel technique for automatically generating the SCOP classification of a protein structure with high accuracy. We achieve accurate classification by combining the decisions of multiple methods using the consensus of a committee (or an ensemble) classifier. Our technique, based on decision trees, is rooted in machine learning which shows that by judicially employing component classifiers, an ensemble classifier can be constructed to outperform its components. We use two sequence- and three structure-comparison tools as component classifiers. Given a protein structure and using the joint hypothesis, we first determine if the protein belongs to an existing category (family, superfamily, fold) in the SCOP hierarchy. For the proteins that are predicted as members of the existing categories, we compute their family-, superfamily-, and fold-level classifications using the consensus classifier. We show that we can significantly improve the classification accuracy compared to the individual component classifiers. In particular, we achieve error rates that are 3-12 times less than the individual classifiers' error rates at the family level, 1.5-4.5 times less at the superfamily level, and 1.1-2.4 times less at the fold level.     

Filling-in Void and Sparse Regions in Protein Sequence Space by Protein-Like Artificial Sequences Enables Remarkable Enhancement in Remote Homology Detection Capability

Protein functional annotation relies on the identification of accurate relationships, sequence divergence being a key factor. This is especially evident when distant protein relationships are demonstrated only with three-dimensional structures. To address this challenge, we describe a computational approach to purposefully bridge gaps between related protein families through directed design of protein-like “linker” sequences. For this, we represented SCOP domain families, integrated with sequence homologues, as multiple profiles and performed HMM-HMM alignments between related domain families. Where convincing alignments were achieved, we applied a roulette wheel-based method to design 3,611,010 protein-like sequences corresponding to 374 SCOP folds. To analyze their ability to link proteins in homology searches, we used 3024 queries to search two databases, one containing only natural sequences and another one additionally containing designed sequences. Our results showed that augmented database searches showed up to 30% improvement in fold coverage for over 74% of the folds, with 52 folds achieving all theoretically possible connections. Although sequences could not be designed between some families, the availability of designed sequences between other families within the fold established the sequence continuum to demonstrate 373 difficult relationships. Ultimately, as a practical and realistic extension, we demonstrate that such protein-like sequences can be “plugged-into” routine and generic sequence database searches to empower not only remote homology detection but also fold recognition. Our richly statistically supported findings show that complementary searches in both databases will increase the effectiveness of sequence-based searches in recognizing all homologues sharing a common fold.