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%.     

Identifying sequence-structure pairs undetected by sequence alignments

We examine how effectively simple potential functions previously developed can identify compatibilities between sequences and structures of proteins for database searches. The potential function consists of pairwise contact energies, repulsive packing potentials of residues for overly dense arrangement and short-range potentials for secondary structures, all of which were estimated from statistical preferences observed in known protein structures. Each potential energy term was modified to represent compatibilities between sequences and structures for globular proteins. Pairwise contact interactions in a sequence-structure alignment are evaluated in a mean field approximation on the basis of probabilities of site pairs to be aligned. Gap penalties are assumed to be proportional to the number of contacts at each residue position, and as a result gaps will be more frequently placed on protein surfaces than in cores. In addition to minimum energy alignments, we use probability alignments made by successively aligning site pairs in order by pairwise alignment probabilities. The results show that the present energy function and alignment method can detect well both folds compatible with a given sequence and, inversely, sequences compatible with a given fold, and yield mostly similar alignments for these two types of sequence and structure pairs. Probability alignments consisting of most reliable site pairs only can yield extremely small root mean square deviations, and including less reliable pairs increases the deviations. Also, it is observed that secondary structure potentials are usefully complementary to yield improved alignments with this method. Remarkably, by this method some individual sequence-structure pairs are detected having only 5-20% sequence identity.     

Comparing protein sequence-based and predicted secondary structure-based methods for identification of remote homologs.

We have compared a novel sequence-structure matching technique, FORESST, for detecting remote homologs to three existing sequence based methods, including local amino acid sequence similarity by BLASTP, hidden Markov models (HMMs) of sequences of protein families using SAM, HMMs based on sequence motifs identified using meta-MEME. FORESST compares predicted secondary structures to a library of structural families of proteins, using HMMs. Altogether 45 proteins from nine structural families in the database CATH were used in a cross-validated test of the fold assignment accuracy of each method. Local sequence similarity of a query sequence to a protein family is measured by the highest segment pair (HSP) score. Each of the HMM-based approaches (FORESST, MEME, amino acid sequence-based HMM) yielded log-odds score for the query sequence. In order to make a fair comparison among these methods, the scores for each method were converted to Z-scores in a uniform way by comparing the raw scores of a query protein with the corresponding scores for a set of unrelated proteins. Z-Scores were analyzed as a function of the maximum pairwise sequence identity (MPSID) of the query sequence to sequences used in training the model. For MPSID above 20%, the Z-scores increase linearly with MPSID for the sequence-based methods but remain roughly constant for FORESST. Below 15%, average Z-scores are close to zero for the sequence-based methods, whereas the FORESST method yielded average Z-scores of 1.8 and 1.1, using observed and predicted secondary structures, respectively. This demonstrates the advantage of the sequence-structure method for detecting remote homologs.     

An integrated approach to the analysis and modeling of protein sequences and structures. II. On the relationship between sequence and structural similarity for proteins that are not obviously related in sequence

Here, we discuss the relationship between protein sequence and protein structural similarity. It is established that a protein structural distance (PSD) of 2.0 is a threshold above which two proteins are unlikely to have a detectable pairwise sequence relationship. A precise correlation is established between the level of sequence similarity, defined by a normalized Smith-Waterman score, and the probability that two proteins will have a similar structure (defined by pairwise PSD<2). This correlation can be used in evaluating the likelihood for success in a comparative modeling procedure. We establish the existence of a correlation between sequence and structural similarity for pairs of proteins that are related in structure but whose sequence relationship is not detectable using standard pairwise sequence alignments. Although it is well known that there is a close relationship between sequence and structural similarity for pairwise sequence identities greater than about 30 %, there has been little discussion as to the possible existence of such a relationship for pairs of proteins in or below the twilight zone of sequence similarity (<25 % pairwise sequence identity). Possible implications of our results for the evolution of protein structure are discussed.     

Persistently conserved positions in structurally similar, sequence dissimilar proteins: Roles in preserving protein fold and function

Many protein pairs that share the same fold do not have any detectable sequence similarity, providing a valuable source of information for studying sequence-structure relationship. In this study, we use a stringent data set of structurally similar, sequence-dissimilar protein pairs to characterize residues that may play a role in the determination of protein structure and/or function. For each protein in the database, we identify amino-acid positions that show residue conservation within both close and distant family members. These positions are termed "persistently conserved". We then proceed to determine the "mutually" persistently conserved (MPC) positions: those structurally aligned positions in a protein pair that are persistently conserved in both pair mates. Because of their intra- and interfamily conservation, these positions are good candidates for determining protein fold and function. We find that 45% of the persistently conserved positions are mutually conserved. A significant fraction of them are located in critical positions for secondary structure determination, they are mostly buried, and many of them form spatial clusters within their protein structures. A substitution matrix based on the subset of MPC positions shows two distinct characteristics: (i) it is different from other available matrices, even those that are derived from structural alignments; (ii) its relative entropy is high, emphasizing the special residue restrictions imposed on these positions. Such a substitution matrix should be valuable for protein design experiments.     

Non-sequential structure-based alignments reveal topology-independent core packing arrangements in proteins

MOTIVATION: Proteins of the same class often share a secondary structure packing arrangement but differ in how the secondary structure units are ordered in the sequence. We find that proteins that share a common core also share local sequence-structure similarities, and these can be exploited to align structures with different topologies. In this study, segments from a library of local sequence-structure alignments were assembled hierarchically, enforcing the compactness and conserved inter-residue contacts but not sequential ordering. Previous structure-based alignment methods often ignore sequence similarity, local structural equivalence and compactness. RESULTS: The new program, SCALI (Structural Core ALIgnment), can efficiently find conserved packing arrangements, even if they are non-sequentially ordered in space. SCALI alignments conserve remote sequence similarity and contain fewer alignment errors. Clustering of our pairwise non-sequential alignments shows that recurrent packing arrangements exist in topologically different structures. For example, the three-layer sandwich domain architecture may be divided into four structural subclasses based on internal packing arrangements. These subclasses represent an intermediate level of structure classification, more general than topology, but more specific than architecture as defined in CATH. A strategy is presented for developing a set of predictive hidden Markov models based on multiple SCALI alignments.     

Conservation of inter-residue interactions and prediction of folding rates of domain repeats

Domains are the main structural and functional units of larger proteins. They tend to be contiguous in primary structure and can fold and function independently. It has been observed that 10–20% of all encoded proteins contain duplicated domains and the average pairwise sequence identity between them is usually low. In the present study, we have analyzed the structural similarity between domain repeats of proteins with known structures available in the Protein Data Bank using structure-based inter-residue interaction measures such as the number of long-range contacts, surrounding hydrophobicity, and pairwise interaction energy. We used RADAR program for detecting the repeats in a protein sequence which were further validated using Pfam domain assignments. The sequence identity between the repeats in domains ranges from 20 to 40% and their secondary structural elements are well conserved. The number of long-range contacts, surrounding hydrophobicity calculations and pairwise interaction energy of the domain repeats clearly reveal the conservation of 3-D structure environment in the repeats of domains. The proportions of mainchain–mainchain hydrogen bonds and hydrophobic interactions are also highly conserved between the repeats. The present study has suggested that the computation of these structure-based parameters will give better clues about the tertiary environment of the repeats in domains. The folding rates of individual domains in the repeats predicted using the long-range order parameter indicate that the predicted folding rates correlate well with most of the experimentally observed folding rates for the analyzed independently folded domains.     

Sequence-structure specificity of a knowledge based energy function at the secondary structure level

MOTIVATION: This paper investigates the sequence-structure specificity of a representative knowledge based energy function by applying it to threading at the level of secondary structures of proteins. Assessing the strengths and weaknesses of an energy function at this fundamental level provides more detailed and insightful information than at the tertiary structure level and the results obtained can be useful in tertiary level threading. RESULTS: We threaded each of the 293 non-redundant proteins onto the secondary structures contained in its respective native protein (host template). We also used 68 pairs of proteins with similar folds and low sequence identity. For each pair, we threaded the sequence of one protein onto the secondary structures of the other protein. The discerning power of the total energy function and its one-body, pairwise, and mutation components is studied. We then applied our energy function to a recent study which demonstrated how a designed 11-amino acid sequence can replace distinct segments (one segment is an alpha-helix, the other is a beta-sheet) of a protein without changing its fold. We conducted random mutations of the designed sequence to determine the patterns for favorable mutations. We also studied the sequence-structure specificity at the boundaries of a secondary structure. Finally, we demonstrated how to speed up tertiary level threading by filtering out alignments found to be energetically unfavorable during the secondary structure threading. AVAILABILITY: The program is available on request from the authors. CONTACT: xud@ornl.gov     

fRMSDPred: predicting local RMSD between structural fragments using sequence information

The effectiveness of comparative modeling approaches for protein structure prediction can be substantially improved by incorporating predicted structural information in the initial sequence-structure alignment. Motivated by the approaches used to align protein structures, this article focuses on developing machine learning approaches for estimating the RMSD value of a pair of protein fragments. These estimated fragment-level RMSD values can be used to construct the alignment, assess the quality of an alignment, and identify high-quality alignment segments. We present algorithms to solve this fragment-level RMSD prediction problem using a supervised learning framework based on support vector regression and classification that incorporates protein profiles, predicted secondary structure, effective information encoding schemes, and novel second-order pairwise exponential kernel functions. Our comprehensive empirical study shows superior results compared with the profile-to-profile scoring schemes. We also show that for protein pairs with low sequence similarity (less than 12% sequence identity) these new local structural features alone or in conjunction with profile-based information lead to alignments that are considerably accurate than those obtained by schemes that use only profile and/or predicted secondary structure information.     

Evolutionary constraints on structural similarity in orthologs and paralogs

Although a quantitative relationship between sequence similarity and structural similarity has long been established, little is known about the impact of orthology on the relationship between protein sequence and structure. Among homologs, orthologs (derived by speciation) more frequently have similar functions than paralogs (derived by duplication). Here, we hypothesize that an orthologous pair will tend to exhibit greater structural similarity than a paralogous pair at the same level of sequence similarity. To test this hypothesis, we used 284,459 pairwise structure‐based alignments of 12,634 unique domains from SCOP as well as orthology and paralogy assignments from OrthoMCL DB. We divided the comparisons by sequence identity and determined whether the sequence‐structure relationship differed between the orthologs and paralogs. We found that at levels of sequence identity between 30 and 70%, orthologous domain pairs indeed tend to be significantly more structurally similar than paralogous pairs at the same level of sequence identity. An even larger difference is found when comparing ligand binding residues instead of whole domains. These differences between orthologs and paralogs are expected to be useful for selecting template structures in comparative modeling and target proteins in structural genomics.     

QUASAR--scoring and ranking of sequence-structure alignments

SUMMARY: Sequence-structure alignments are a common means for protein structure prediction in the fields of fold recognition and homology modeling, and there is a broad variety of programs that provide such alignments based on sequence similarity, secondary structure or contact potentials. Nevertheless, finding the best sequence-structure alignment in a pool of alignments remains a difficult problem. QUASAR (quality of sequence-structure alignments ranking) provides a unifying framework for scoring sequence-structure alignments that aids finding well-performing combinations of well-known and custom-made scoring schemes. Those scoring functions can be benchmarked against widely accepted quality scores like MaxSub, TMScore, Touch and APDB, thus enabling users to test their own alignment scores against 'standard-of-truth' structure-based scores. Furthermore, individual score combinations can be optimized with respect to benchmark sets based on known structural relationships using QUASAR's in-built optimization routines.     

MARNA: multiple alignment and consensus structure prediction of RNAs based on sequence structure comparisons

MOTIVATION: Due to the importance of considering secondary structures in aligning functional RNAs, several pairwise sequence-structure alignment methods have been developed. They use extended alignment scores that evaluate secondary structure information in addition to sequence information. However, two problems for the multiple alignment step remain. First, how to combine pairwise sequence-structure alignments into a multiple alignment and second, how to generate secondary structure information for sequences whose explicit structural information is missing. RESULTS: We describe a novel approach for multiple alignment of RNAs (MARNA) taking into consideration both the primary and the secondary structures. It is based on pairwise sequence-structure comparisons of RNAs. From these sequence-structure alignments, libraries of weighted alignment edges are generated. The weights reflect the sequential and structural conservation. For sequences whose secondary structures are missing, the libraries are generated by sampling low energy conformations. The libraries are then processed by the T-Coffee system, which is a consistency based multiple alignment method. Furthermore, we are able to extract a consensus-sequence and -structure from a multiple alignment. We have successfully tested MARNA on several datasets taken from the Rfam database.     

Protein-segment universe exhibiting transitions at intermediate segment length in conformational subspaces

Background

Lipocalins are widely distributed in nature and are found in bacteria, plants, arthropoda and vertebra. In hematophagous arthropods, they are implicated in the successful accomplishment of the blood meal, interfering with platelet aggregation, blood coagulation and inflammation and in the transmission of disease parasites such as Trypanosoma cruzi and Borrelia burgdorferi. The pairwise sequence identity is low among this family, often below 30%, despite a well conserved tertiary structure. Under the 30% identity threshold, alignment methods do not correctly assign and align proteins. The only safe way to assign a sequence to that family is by experimental determination. However, these procedures are long and costly and cannot always be applied. A way to circumvent the experimental approach is sequence and structure analyze. To further help in that task, the residues implicated in the stabilisation of the lipocalin fold were determined. This was done by analyzing the conserved interactions for ten lipocalins having a maximum pairwise identity of 28% and various functions.

Results

It was determined that two hydrophobic clusters of residues are conserved by analysing the ten lipocalin structures and sequences. One cluster is internal to the barrel, involving all strands and the 3₁₀ helix. The other is external, involving four strands and the helix lying parallel to the barrel surface. These clusters are also present in RaHBP2, a unusual "outlier" lipocalin from tick Rhipicephalus appendiculatus. This information was used to assess assignment of LIR2 a protein from Ixodes ricinus and to build a 3D model that helps to predict function. FTIR data support the lipocalin fold for this protein.

Conclusion

By sequence and structural analyzes, two conserved clusters of hydrophobic residues in interactions have been identified in lipocalins. Since the residues implicated are not conserved for function, they should provide the minimal subset necessary to confer the lipocalin fold. This information has been used to assign LIR2 to lipocalins and to investigate its structure/function relationship. This study could be applied to other protein families with low pairwise similarity, such as the structurally related fatty acid binding proteins or avidins.     

Quantification of tertiary structural conservation despite primary sequence drift in the globin fold.

The globin family of protein structures was the first for which it was recognized that tertiary structure can be highly conserved even when primary sequences have diverged to a virtually undetectable level of similarity. This principle of structural inertia in molecular evolution is now evident for many other protein families. We have performed a systematic comparison of the sequences and structures of 6 representative hemoglobin subunits as diverse in origin as plants, clams, and humans. Our analysis is based on a 97-residue helical core in common to all 6 structures. Amino acid sequence identities range from 12.4% to 42.3% in pairwise comparisons, and, despite these variations, the maximal RMS deviation in alpha-carbon positions is 3.02 A. Overall, sequence similarity and structural deviation are significantly anticorrelated, with a correlation coefficient of -0.71, but for a set of structures having under 20% pairwise identity, this anticorrelation falls to -0.38, which emphasizes the weak connection between a specific sequence and the tertiary fold. There is substantial variability in structure outside the helical core, and functional characteristics of these globins also differ appreciably. Nevertheless, despite variations in detail that the sequence dissimilarities and functional differences imply, the core structures of these globins remain remarkably preserved.     

MANIFOLD: protein fold recognition based on secondary structure, sequence similarity and enzyme classification

We present a protein fold recognition method, MANIFOLD, which uses the similarity between target and template proteins in predicted secondary structure, sequence and enzyme code to predict the fold of the target protein. We developed a non-linear ranking scheme in order to combine the scores of the three different similarity measures used. For a difficult test set of proteins with very little sequence similarity, the program predicts the fold class correctly in 34% of cases. This is an over twofold increase in accuracy compared with sequence-based methods such as PSI-BLAST or GenTHREADER, which score 13-14% correct first hits for the same test set. The functional similarity term increases the prediction accuracy by up to 3% compared with using the combination of secondary structure similarity and PSI-BLAST alone. We argue that using functional and secondary structure information can increase the fold recognition beyond sequence similarity.     

Evaluation of threading specificity and accuracy

Threading experiments with proteins from the globin family provide an indication of the nature of the structural similarity required for successful fold recognition and accurate sequence-structure alignment. Threading scores are found to rise above the noise of false positives whenever roughly 60% of residues from a sequence can be aligned with analogous sites in the structure of a remote homolog. Fold recognition specificity thus appears to be limited by the extent of structural similarity, regardless of the degree of sequence similarity. Threading alignment accuracy is found to depend more critically on the degree of structural similarity. Alignments are accurate, placing the majority of residues exactly as in structural alignment, only when superposition residuals are less than 2.5 Å. These criteria for successful recognition and sequence-structure alignment appear to be consistent with the successes and failures of threading methods in blind structure prediction. They also suggest a direct assay for improved threading methods: Potentials and alignment models should be tested for their ability to detect less extensive structural similarities, and to produce accurate alignments when superposition residuals for this conserved “core” fall in the range characteristic of remote homologs. © 1996 Wiley-Liss, Inc.

1 This article is a US Government work and, as such, is in the public domain in the United States of America.

     

Protein homology detection and fold inference through multiple alignment entropy profiles

Homology detection and protein structure prediction are central themes in bioinformatics. Establishment of relationship between protein sequences or prediction of their structure by sequence comparison methods finds limitations when there is low sequence similarity. Recent works demonstrate that the use of profiles improves homology detection and protein structure prediction. Profiles can be inferred from protein multiple alignments using different approaches. The "Conservatism-of-Conservatism" is an effective profile analysis method to identify structural features between proteins having the same fold but no detectable sequence similarity. The information obtained from protein multiple alignments varies according to the amino acid classification employed to calculate the profile. In this work, we calculated entropy profiles from PSI-BLAST-derived multiple alignments and used different amino acid classifications summarizing almost 500 different attributes. These entropy profiles were converted into pseudocodes which were compared using the FASTA program with an ad-hoc matrix. We tested the performance of our method to identify relationships between proteins with similar fold using a nonredundant subset of sequences having less than 40% of identity. We then compared our results using Coverage Versus Error per query curves, to those obtained by methods like PSI-BLAST, COMPASS and HHSEARCH. Our method, named HIP (Homology Identification with Profiles) presented higher accuracy detecting relationships between proteins with the same fold. The use of different amino acid classifications reflecting a large number of amino acid attributes, improved the recognition of distantly related folds. We propose the use of pseudocodes representing profile information as a fast and powerful tool for homology detection, fold assignment and analysis of evolutionary information enclosed in protein profiles.     

Residue-residue contact substitution probabilities derived from aligned three-dimensional structures and the identification of common folds.

We report the derivation of scores that are based on the analysis of residue-residue contact matrices from 443 3-dimensional structures aligned structurally as 96 families, which can be used to evaluate sequence-structure matches. Residue-residue contacts and the more than 3 x 10(6) amino acid substitutions that take place between pairs of these contacts at aligned positions within each family of structures have been tabulated and segregated according to the solvent accessibility of the residues involved. Contact maps within a family of structures are shown to be highly conserved (approximately 75%) even when the sequence identity is approaching 10%. In a comparison involving a globin structure and the search of a sequence databank (> 21,000 sequences), the contact probability scores are shown to provide a very powerful secondary screen for the top scoring sequence-structure matches, where between 69% and 84% of the unrelated matches are eliminated. The search of an aligned set of 2 globins against a sequence databank and the subsequent residue contact-based evaluation of matches locates all 618 globin sequences before the first non-globin match. From a single bacterial serine proteinase structure, the structural template approach coupled with residue-residue contact substitution data lead to the detection of the mammalian serine proteinase family among the top matches in the search of a sequence databank.     

How well is enzyme function conserved as a function of pairwise sequence identity?

Enzyme function conservation has been used to derive the threshold of sequence identity necessary to transfer function from a protein of known function to an unknown protein. Using pairwise sequence comparison, several studies suggested that when the sequence identity is above 40%, enzyme function is well conserved. In contrast, Rost argued that because of database bias, the results from such simple pairwise comparisons might be misleading. Thus, by grouping enzyme sequences into families based on sequence similarity and selecting representative sequences for comparison, he showed that enzyme function starts to diverge quickly when the sequence identity is below 70%. Here, we employ a strategy similar to Rost's to reduce the database bias; however, we classify enzyme families based not only on sequence similarity, but also on functional similarity, i.e. sequences in each family must have the same four digits or the same first three digits of the enzyme commission (EC) number. Furthermore, instead of selecting representative sequences for comparison, we calculate the function conservation of each enzyme family and then average the degree of enzyme function conservation across all enzyme families. Our analysis suggests that for functional transferability, 40% sequence identity can still be used as a confident threshold to transfer the first three digits of an EC number; however, to transfer all four digits of an EC number, above 60% sequence identity is needed to have at least 90% accuracy. Moreover, when PSI-BLAST is used, the magnitude of the E-value is found to be weakly correlated with the extent of enzyme function conservation in the third iteration of PSI-BLAST. As a result, functional annotation based on the E-values from PSI-BLAST should be used with caution. We also show that by employing an enzyme family-specific sequence identity threshold above which 100% functional conservation is required, functional inference of unknown sequences can be accurately accomplished. However, this comes at a cost: those true positive sequences below this threshold cannot be uniquely identified.