Automatic definition of recurrent local structure motifs in proteins.

An automatic procedure for defining recurrent folding motifs in proteins of known structure is described. These motifs are formed by short polypeptide fragments of equal size containing between four and seven residues. The method applies a classical clustering algorithm that operates on distances between selected backbone atoms. In one application, we use it to cluster all protein fragments into only four structural classes. This classification is rough considering the observed diversity of local structures, but comparable in homogeneity to the four classes of secondary structure (alpha-helix, beta-strand, turn and coil). Yet, it discriminates between extended and curved coil and distinguishes beta-bulges from beta-strands. In a second application, the clustering procedure is combined with assignment of backbone dihedral angles to allowed regions in the Ramachandran map. This produces an exhaustive repertoire of highly homogeneous families of structural motifs that contains all the beta-hairpins, beta alpha- and alpha beta-loops previously defined by manual procedures, and new structural families of which two examples, a beta alpha-loop and an alpha-helix beginning, are analyzed in detail. The described automatic procedures should be useful in categorizing structure information in proteins, thereby increasing our ability to analyze relations between structure and sequence.

Relations between protein sequence and structure and their significance.

The relation between amino acid sequence and local structure in proteins is investigated. The local structures considered are either the four classes of secondary structure (H, E, T and C) or four classes of local conformations defined using measures of conformational similarity based on distances between C alpha atoms. The classes are obtained by applying an automatic clustering procedure to short polypeptide fragments of uniform length from a database of 75 known protein structures. The thrust of our investigation consists of systematically searching the database for simple amino acid patterns of the type Gly-X-Ala-X-X-Val, where X denotes an arbitrary residue. Patterns that are nearly always associated with the same structure are retained. Finding many such associations, we then evaluate by a statistical approach how many among them are non-random and compare the results for different definitions of local structure. A similar comparison is made for the predictive value of retained associations, which is assessed using an internal test based on dividing the database into "learning" and "test" subsets. While we find that local structures defined by conformational similarity are not superior to secondary structure for prediction purposes, they help us gain insight into the factors that influence the predictive value of derived associations. A major conclusion is that the number of retained associations is in large excess over the number expected from a random correlation between sequence and structure, irrespective of how local conformation is defined. However, only a very small number of these associations can be earmarked as reliable using statistical criteria, due to the limited size of the database. We find, for instance, that the pattern Ala-Ala-X-X-Lys reliably characterizes helix, and the pattern Val-X-Val-X-X-X-Ala reliably characterizes extended structure and beta-strand. The possibility is discussed that these and other reliable associations correspond to regions of the polypeptide chain whose conformations are locally determined and that these regions may play a role in folding.

Prediction of protein backbone conformation based on seven structure assignments. Influence of local interactions.

A method is developed to compute backbone tertiary folds from the amino acid sequence. In this method, the number of degrees of freedom is drastically reduced by neglecting side-chain flexibility, and by describing backbone conformations as combinations of only seven structural states. These are characterized by single values of the dihedral angles phi, psi and omega, representing allowed conformations of the isolated dipeptide. We show that this restrictive model is none the less capable of describing native backbones to within acceptable deviations. Using our backbone description, potentials of mean force are derived from a database of known protein structures, based on statistical influences of single residues and residue pairs on the conformational states in their vicinity along the chain. This yields the force-field component due to local interactions, which is then used to predict lowest-energy conformations from any given amino acid sequence. The prediction algorithm does not require searching conformational space and is therefore extremely fast. Another important asset of our method is that it is able to compute not only the minimum energy conformation, but any number of lowest energy structures, whose relative preferences can be determined from the corresponding computed energy values. The performance of our procedure is tested on short peptides that are likely to be stabilized by local interactions. These include several helical structures and a hexapeptide with a beta-bend conformation, corresponding to peptides shown to have relatively well-defined conformations in aqueous solution, and to protein segments believed to adopt their native conformation early during folding. In addition, several flexible peptides are analysed. Except for the problems encountered in predicting observed disulphide bridges in two of the flexible peptides, and in a somewhat larger fragment comprising residues 30 to 51 of bovine trypsin inhibitor, prediction results compare very favourably with experimental data. Potential applications of our procedure to protein modelling and its extension to protein folding are discussed.

Automatic classification and analysis of alpha alpha-turn motifs in proteins.

An automatic procedure for the classification of short protein fragments, representing turn motifs between two consecutive secondary structures, is presented. This procedure has two steps. Fragments of given length are first grouped on the basis of their backbone dihedral angle values, and then clustered as a function of the root-mean-square deviation of their superimposed backbone atoms. The classification procedure identifies 63 families of turn motifs with at least five members, in a dataset of 141 proteins. A detailed analysis is presented of the ten identified alpha alpha-turn families, of which four correspond to novel motifs. The sequence and structure features that characterize these families are described. It is found that some features are conserved within the fragments belonging to the same family, but their environment in the parent protein varies considerably. N-capping interactions and helix stop signals are encountered in a number of families, where they seem to stabilize the motif conformation. In two families, one with three residues in the loop, and one with four, an appreciable fraction of the members displays both types of characteristic helix end interactions in the same motif. Interestingly, contrary to most other alpha alpha-turns, the relative frequency of these two motifs is much higher than that of short protein segments with the same loop conformation. Furthermore, the family with three residues in the loop includes the helix-turn-helix motif known to bind DNA. It seems to be the only one among the ten identified families that can be related to biological function.

Factors influencing the ability of knowledge-based potentials to identify native sequence-structure matches.

Several types of potentials are derived from a dataset of known protein structures by computing statistical relations between amino acid sequence and different descriptions of the protein conformation. These potentials formulate in different ways backbone dihedral angle preferences, pairwise distance-dependent interactions between amino acid residues, and solvation effects based on accessible surface area calculations. Parameters affecting the characteristics and the performance of the potentials are critically assessed by monitoring recognition of the native fold in a strict screening test, where each sequence in the dataset is threaded through a repertoire of motifs, generated from all corresponding structures. Sequence gaps are not allowed, to avoid additional approximations. Results show that residue interaction potentials computed from distances between average side-chain centroids perform significantly better on this test than those computed considering inter-C alpha or inter-C beta distances. Combining potentials that are based on different structural descriptions and different interactions is also beneficial. The performance of some of these potentials is in fact so good that they recognize the correct fold for all the tested proteins, including subunits known to be unstable in the absence of quaternary interactions. Most strikingly, potentials representing backbone dihedral angle preferences recognize as many as 68 protein chains out of a total of 74, even though they consider solely local interactions along the chain, which, being the same as those considered in secondary structure prediction methods, are well known to be incapable of determining the full three-dimensional fold. This leads us to question the ability of procedures that screen a limited repertoire of structures to act as a stringent test for the potentials. We concede, however, that they are useful and fast tests, capable of revealing gross shortcomings of the potentials, or possible biases towards native recognition due, for example, to effects of sequence memory.

Automatic analysis of protein conformational changes by multiple linkage clustering.

An automatic algorithm is presented for analyzing protein conformational changes such as those occurring upon substrate binding or in different crystal forms of the same protein. Using, as sole information, the atomic coordinates of a pair of protein structures, the procedure first generates structure alignments, which optimize the root-mean-square deviation of the backbone atoms. To this end, equivalent secondary structures and/or loops from both proteins are combined by a multiple linkage hierarchic clustering algorithm, which generates several intertwined clustering trees. Automatic analysis of these clustering trees is used to dissect the mechanism of the conformational change. It allows the identification of the static core, representing the collection of secondary structures which undergo no structural changes, as well as other entities which move like rigid bodies. It also permits the description of the movement of secondary structures or loops relative to this core or entities. USing this information, it can be inferred whether a particular conformational change involves shear or hinge motion, or components of both. The algorithm is applied to the analysis of the conformational changes of citrate synthase, lactate dehydrogenase, lactoferrin and beta-glucosyltransferase, representing typical examples of shear- and hinge-type mechanisms, and a varied range in movement size. The results are shown to be in excellent agreement with previous analyses, and to provide additional information which gives a more complete and objective picture of the conformational change. Using our automatic algorithm, we find that any conformational change may be viewed as having components of both shear- and hinge-type motion. Determining which of these is most appropriate requires the combination of the information provided by our procedure with detailed knowledge of the protein tertiary structures.

Weak correlation between predictive power of individual sequence patterns and overall prediction accuracy in proteins.

Patterns in amino acid properties (polar, hydrophobic, etc.) that characterize secondary structure motifs are derived from a database containing 75 protein structures, with the aim of circumventing the limitations due to data base size so as to increase structure prediction score. Many such sequence-structure associations with high intrinsic predictive power are found, which turn out to be correct 78% of the time when applied individually to proteins outside the learning set. Based on these associations, a prediction method is developed, which reaches the score of 62% on the 3 states alpha-helix, beta-strand, and loop, without using additional constraints. Though this score is quite good compared to that of other available prediction methods, it is much lower than could be expected from the high intrinsic predictive power of the associations used. The reasons underlying this surprising result, which indicate that prediction score and intrinsic predictive power are only weakly coupled, are discussed. It is also shown that the size of the present database still seriously limits prediction scores, even when property patterns are used, and that higher scores are expected in large databases. Clues are provided on the relative influence of neglecting spatial interactions on prediction efficiency, suggesting that, in sufficiently large databases, predicted secondary structures would correspond to those formed early in the folding process. This hypothesis is tested by confronting present predictions with available experimental data on early protein folding intermediates and on small peptides that adopt a relatively stable conformation in water. Although admittedly there are still too few such data, results suggest that the hypothesis might be well founded.

Extracting information on folding from the amino acid sequence: consensus regions with preferred conformation in homologous proteins.

It is investigated whether protein segments predicted to have a well-defined conformational preference in the absence of tertiary interactions are conserved in families of homologous proteins. The prediction method follows the procedures of Rooman, M., Kocher, J.-P., and Wodak, S. (preceding paper in this issue). It uses a knowledge-based force field that incorporates only local interactions along the sequence and identifies segments whose lowest energy structure displays a sizable energy gap relative to other computed conformations. In 13 of the protein families and subfamilies considered that are sufficiently homologous to have similar 3D structures, at least one region is consistently predicted as having the same preferred conformation in virtually all family members. These regions are between 4 and 26 residues long. They are often located at chain ends and correspond primarily to segments of secondary structure heavily involved in interactions with the rest of the protein, suggesting that they could act as nuclei around which other parts of the structure would assemble. Experimental data on early folding intermediates or on protein fragments with appreciable structure in aqueous solution are available for more than half of the protein families. Comparison of our results with these data is quite favorable. They reveal that each of the experimentally identified early formed, or independently stable, substructures harbors at least one of the segments consistently predicted as having a preferred conformation by our procedure. The implications of our findings for the conservation of folding pathways in homologous proteins are discussed.

Are database-derived potentials valid for scoring both forward and inverted protein folding?

Database-derived potentials, compiled from frequencies of sequence and structure features, are often used for scoring the compatibility of protein sequences and conformations. It is often believed that these scores correspond to differences in free energy with, in addition, a term containing the partition function of the system. Since this function does not depend on the conformation, the potentials are considered to be valid for scoring the compatibility of different conformations with a given sequence ('forward folding'), but not of sequences with a given structure ('inverted folding'). This interpretation is questioned here. It is argued that when many body-effects, which dominate frequencies compiled from the protein database, are corrected for, the potentials approximate a physically meaningful free energy difference from which the partition function term cancels out. It is the difference between the free energy of a given sequence in a specific conformation and that of the same sequence in a denatured-like state. Two examples of denatured-like states are discussed. Depending on the considered state, the free energy difference reduces to the commonly used scoring scheme, or contains additional terms that depend on the sequence. In both cases, all the terms can be derived from sequence-structure frequencies in the database. Such free energy difference, commonly defined as the folding free energy, is a measure of protein stability and can be used for scoring both forward and inverted protein folding. The implications for the use of knowledge-based potentials in protein structure prediction are described. Finally, the difficulty of designing tests that could validate the proposed approach, and the inherent limitations of such tests, are discussed.

Identification of predictive sequence motifs limited by protein structure data base size.

Associations between short amino acid sequence patterns and protein secondary structure classes can be found by searching a data base of known protein structures. Analysis of these associations suggests that secondary structure of proteins can be determined locally by sequence motifs of high predictive value, but at present our ability to find these motifs is limited by the size of the available data bases.

Protein structure prediction by threading methods: evaluation of current techniques.

This paper evaluates the results of a protein structure prediction contest. The predictions were made using threading procedures, which employ techniques for aligning sequences with 3D structures to select the correct fold of a given sequence from a set of alternatives. Nine different teams submitted 86 predictions, on a total of 21 target proteins with little or no sequence homology to proteins of known structure. The 3D structures of these proteins were newly determined by experimental methods, but not yet published or otherwise available to the predictors. The predictions, made from the amino acid sequence alone, thus represent a genuine test of the current performance of threading methods. Only a subset of all the predictions is evaluated here. It corresponds to the 44 predictions submitted for the 11 target proteins seen to adopt known folds. The predictions for the remaining 10 proteins were not analyzed, although weak similarities with known folds may also exist in these proteins. We find that threading methods are capable of identifying the correct fold in many cases, but not reliably enough as yet. Every team predicts correctly a different set of targets, with virtually all targets predicted correctly by at least one team. Also, common folds such as TIM barrels are recognized more readily than folds with only a few known examples. However, quite surprisingly, the quality of the sequence-structure alignments, corresponding to correctly recognized folds, is generally very poor, as judged by comparison with the corresponding 3D structure alignments. Thus, threading can presently not be relied upon to derive a detailed 3D model from the amino acid sequence. This raises a very intriguing question: how is fold recognition achieved? Our analysis suggests that it may be achieved because threading procedures maximize hydrophobic interactions in the protein core, and are reasonably good at recognizing local secondary structure.

Structural classification of alphabetabeta and betabetaalpha supersecondary structure units in proteins.

We present a fully automatic structural classification of supersecondary structure units, consisting of two hydrogen-bonded beta strands, preceded or followed by an alpha helix. The classification is performed on the spatial arrangement of the secondary structure elements, irrespective of the length and conformation of the intervening loops. The similarity of the arrangements is estimated by a structure alignment procedure that uses as similarity measure the root mean square deviation of superimposed backbone atoms. Applied to a set of 141 well-resolved nonhomologous protein structures, the classification yields 11 families of recurrent arrangements. In addition, fragments that are structurally intermediate between the families are found; they reveal the continuity of the classification. The analysis of the families shows that the alpha helix and beta hairpin axes can adopt virtually all relative orientations, with, however, some preferable orientations; moreover, according to the orientation, preferences in the left/right handedness of the alpha-beta connection are observed. These preferences can be explained by favorable side by side packing of the alpha helix and the beta hairpin, local interactions in the region of the alpha-beta connection or stabilizing environments in the parent protein. Furthermore, fold recognition procedures and structure prediction algorithms coupled to database-derived potentials suggest that the preferable nature of these arrangements does not imply their intrinsic stability. They usually accommodate a large number of sequences, of which only a subset is predicted to stabilize the motif. The motifs predicted as stable could correspond to nuclei formed at the very beginning of the folding process.

Amino acid sequence templates derived from recurrent turn motifs in proteins: critical evaluation of their predictive power.

Amino acid sequence patterns suggested to characterize specific recurrent turn conformation in protein are tested as to their predictive power in a database containing 75 proteins of known structure. Many of these patterns are found to be associated with local structures that differ from the motifs originally used to derive them. It is therefore concluded that, while they could be useful for improving predictions made by other methods, their stand-alone predictive power is poor. The issue of deriving and validating consensus sequence patterns for use in protein structure prediction is raised.

Extracting information on folding from the amino acid sequence: accurate predictions for protein regions with preferred conformation in the absence of tertiary interactions.

A recently developed procedure to predict backbone structure from the amino acid sequence [Rooman, M., Kocher, J. P., & Wodak, S. (1991) J. Mol. Biol, 221, 961-979] is fine tuned to identify protein segments, of length 5-15 residues, that adopt well-defined conformations in the absence of tertiary interactions. These segments are obtained by requiring that their predicted lowest energy structures have a sizable energy gap relative to other computed conformations. Applying this procedure to 69 proteins of known structure, we find that regions with largest energy gaps--those having highly preferred conformations--are also the most accurately predicted ones. On the basis of previous findings that such regions correlate well with sites that become structured early during folding, our approach provides the means of identifying such sites in proteins without prior knowledge of the tertiary structure. Furthermore, when predictions are performed so as to ignore the influence of residues flanking each segment along the sequence, a situation akin to excising the considered peptide from the rest of the chain, they offer the possibility of identifying protein segments liable to adopt well-defined conformations on their own. The described approach should have useful applications in experimental and theoretical investigations of protein folding and stability, and aid in designing peptide drugs and vaccines.

Optimal protein structure alignments by multiple linkage clustering: application to distantly related proteins.

A fully automatic procedure for aligning two protein structures is presented. It uses as sole structural similarity measure the root mean square (r.m.s.) deviation of superimposed backbone atoms (N, C alpha, C and O) and is designed to yield optimal solutions with respect to this measure. In a first step, the procedure identifies protein segments with similar conformations in both proteins. In a second step, a novel multiple linkage clustering algorithm is used to identify segment combinations which yield optimal global structure alignments. Several structure alignments can usually be obtained for a given pair of proteins, which are exploited here to define automatically the common structural core of a protein family. Furthermore, an automatic analysis of the clustering trees is described which enables detection of rigid-body movements between structure elements. To illustrate the performance of our procedure, we apply it to families of distantly related proteins. One groups the three alpha + beta proteins ubiquitin, ferredoxin and the B1-domain of protein G. Their common structure motif consists of four beta-strands and the only alpha-helix, with one strand and the helix being displaced as a rigid body relative to the remaining three beta-strands. The other family consists of beta-proteins from the Greek key group, in particular actinoxanthin, the immunoglobulin variable domain and plastocyanin. Their consensus motif, composed of five beta-strands and a turn, is identified, mostly intact, in all Greek key proteins except the trypsins, and interestingly also in three other beta-protein families, the lipocalins, the neuraminidases and the lectins. This result provides new insights into the evolutionary relationships in the very diverse group of all beta-proteins.

Conformational properties of four peptides corresponding to alpha-helical regions of Rhodospirillum cytochrome c2 and bovine calcium binding protein.

Four peptides corresponding to alpha-helical regions delimited by residues 63-73 and 97-112 of cytochrome c2 (Rhodospirillum) and residues 24-36 and 45-55 of bovine calcium binding protein are predicted to be alpha-helical by a recently developed method [Rooman, M., Kocher, J.P., & Wodak, S.J. (1991) J. Mol. Biol. 221, 961-979], synthesized by solid phase methods, and purified by HPLC, and their solution conformations are determined by NMR and CD. The observed conformational properties of these peptides in solution confirmed prediction results: in water/TFE (60/40, v/v) at room temperature, these peptides adopt an alpha-helical conformation, as shown by an extended pattern of strong, sequential dNN(i,i + 1) NOE cross-peaks, d alpha N(i,i + 1) NOEs of reduced intensity, several medium-range [d alpha N(i,i + 3), d alpha N(i,i + 4), d alpha beta-(i,i + 3)] NOE connectivities, small 3JH alpha N values, and more upfield alpha-proton chemical shifts. CD studies at different TFE concentrations and at room temperature provide further evidence of the propensity of these peptides to adopt an alpha-helical conformation in solution, as determined by the ellipticity values at 222 nm, and by deconvolution of the CD spectra. According to the method used, helicities in the range 34-50% and 55-75% are found for the 63-73 and 97-112 fragments of cytochrome c2, respectively, and in the range 53-80% and 42-65% for the fragments 24-36 and 45-55 of calcium binding protein in water/TFE (60/40, v/v) at 298 K. In addition, the experiments and predictions agree for those residues that are more flexible. Finally, the relevance of our results for the protein folding pathways is discussed.