A correlation between amino acid composition and protein structure

Deep learning methods in protein structure prediction

Protein Structure Prediction is a central topic in Structural Bioinformatics. Since the '60s statistical methods, followed by increasingly complex Machine Learning and recently Deep Learning methods, have been employed to predict protein structural information at various levels of detail. In this review, we briefly introduce the problem of protein structure prediction and essential elements of Deep Learning (such as Convolutional Neural Networks, Recurrent Neural Networks and basic feed-forward Neural Networks they are founded on), after which we discuss the evolution of predictive methods for one-dimensional and two-dimensional Protein Structure Annotations, from the simple statistical methods of the early days, to the computationally intensive highly-sophisticated Deep Learning algorithms of the last decade. In the process, we review the growth of the databases these algorithms are based on, and how this has impacted our ability to leverage knowledge about evolution and co-evolution to achieve improved predictions. We conclude this review outlining the current role of Deep Learning techniques within the wider pipelines to predict protein structures and trying to anticipate what challenges and opportunities may arise next.

Deeper Profiles and Cascaded Recurrent and Convolutional Neural Networks for state-of-the-art Protein Secondary Structure Prediction

Protein Secondary Structure prediction has been a central topic of research in Bioinformatics for decades. In spite of this, even the most sophisticated ab initio SS predictors are not able to reach the theoretical limit of three-state prediction accuracy (88–90%), while only a few predict more than the 3 traditional Helix, Strand and Coil classes. In this study we present tests on different models trained both on single sequence and evolutionary profile-based inputs and develop a new state-of-the-art system with Porter 5. Porter 5 is composed of ensembles of cascaded Bidirectional Recurrent Neural Networks and Convolutional Neural Networks, incorporates new input encoding techniques and is trained on a large set of protein structures. Porter 5 achieves 84% accuracy (81% SOV) when tested on 3 classes and 73% accuracy (70% SOV) on 8 classes on a large independent set. In our tests Porter 5 is 2% more accurate than its previous version and outperforms or matches the most recent predictors of secondary structure we tested. When Porter 5 is retrained on SCOPe based sets that eliminate homology between training/testing samples we obtain similar results. Porter is available as a web server and standalone program at http://distilldeep.ucd.ie/porter/ alongside all the datasets and alignments.

Comparison of Peptide Ion Conformers Arising from Non-Helical and Helical Peptides Using Ion Mobility Spectrometry and Gas-Phase Hydrogen/Deuterium Exchange

The dominant gas-phase conformer of [M+3H]³⁺ ions of the model peptide acetyl-PSSSSKSSSSKSSSSKSSSSK has been examined with ion mobility spectrometry (IMS), gas-phase hydrogen deuterium exchange (HDX), and mass spectrometry (MS) techniques. The [M+3H]³⁺ peptide ions are observed predominantly as a relatively compact conformer type. Upon subjecting these ions to electron transfer dissociation (ETD), the level of protection for each amino acid residue in the peptide sequence is assessed. The overall per-residue deuterium uptake is observed to be relatively more efficient for the neutral residues than for the model peptide acetyl-PAAAAKAAAAKAAAAKAAAAK. In comparison, the N-terminal and C-terminal regions of the serine peptide show greater relative protection compared with interior residues. Molecular dynamics (MD) simulations have been used to generate candidate structures for collision cross section and HDX reactivity matching. Hydrogen accessibility scoring (HAS) for select structural candidates from MD simulations has been used to suggest conformer types that could contribute to the observed HDX patterns. The results are discussed with respect to recent studies employing extensive MD simulations of gas-phase structure establishment of a peptide system. Graphical Abstract ᅟ.

Protein stability: a crystallographer's perspective

Protein stability is a topic of major interest for the biotechnology, pharmaceutical and food industries, in addition to being a daily consideration for academic researchers studying proteins. An understanding of protein stability is essential for optimizing the expression, purification, formulation, storage and structural studies of proteins. In this review, discussion will focus on factors affecting protein stability, on a somewhat practical level, particularly from the view of a protein crystallographer. The differences between protein conformational stability and protein compositional stability will be discussed, along with a brief introduction to key methods useful for analyzing protein stability. Finally, tactics for addressing protein-stability issues during protein expression, purification and crystallization will be discussed.

Amino acid pair- and triplet-wise groupings in the interior of α-helical segments in proteins

A statistical approach has been applied to analyse primary structure patterns at inner positions of α-helices in proteins. A systematic survey was carried out in a recent sample of non-redundant proteins selected from the Protein Data Bank, which were used to analyse α-helix structures for amino acid pairing patterns. Only residues more than three positions apart from both termini of the α-helix were considered as inner. Amino acid pairings i, i+k (k=1, 2, 3, 4, 5), were analysed and the corresponding 20×20 matrices of relative global propensities were constructed. An analysis of (i, i+4, i+8) and (i, i+3, i+4) triplet patterns was also performed. These analysis yielded information on a series of amino acid patterns (pairings and triplets) showing either high or low preference for α-helical motifs and suggested a novel approach to protein alphabet reduction. In addition, it has been shown that the individual amino acid propensities are not enough to define the statistical distribution of these patterns. Global pair propensities also depend on the type of pattern, its composition and orientation in the protein sequence. The data presented should prove useful to obtain and refine useful predictive rules which can further the development and fine-tuning of protein structure prediction algorithms and tools.

Using genetic algorithms to select most predictive protein features

Many important characteristics of proteins such as biochemical activity and subcellular localization present a challenge to machine-learning methods: it is often difficult to encode the appropriate input features at the residue level for the purpose of making a prediction for the entire protein. The problem is usually that the biophysics of the connection between a machine-learning method's input (sequence feature) and its output (observed phenomenon to be predicted) remains unknown; in other words, we may only know that a certain protein is an enzyme (output) without knowing which region may contain the active site residues (input). The goal then becomes to dissect a protein into a vast set of sequence-derived features and to correlate those features with the desired output. We introduce a framework that begins with a set of global sequence features and then vastly expands the feature space by generically encoding the coexistence of residue-based features. It is this combination of individual features, that is the step from the fractions of serine and buried (input space 20 + 2) to the fraction of buried serine (input space 20 * 2) that implicitly shifts the search space from global feature inputs to features that can capture very local evidence such as a the individual residues of a catalytic triad. The vast feature space created is explored by a genetic algorithm (GA) paired with neural networks and support vector machines. We find that the GA is critical for selecting combinations of features that are neither too general resulting in poor performance, nor too specific, leading to overtraining. The final framework manages to effectively sample a feature space that is far too large for exhaustive enumeration. We demonstrate the power of the concept by applying it to prediction of protein enzymatic activity.

Amino acid pairing at the N- and C-termini of helical segments in proteins

A systematic survey was carried out in an unbiased sample of 815 protein chains with a maximum of 20% homology selected from the Protein Data Bank, whose structures were solved at a resolution higher than 1.6 A and with a R-factor lower than 25%. A set of 5556 subsequences with alpha-helix or 3(10)-helix motifs was extracted from the protein chains considered. Global and local propensities were then calculated for all possible amino acid pairs of the type (i, i + 1), (i, i + 2), (i, i + 3), and (i, i + 4), starting at the relevant helical positions N1, N2, N3, C3, C2, C1, and N-int (interior positions), and also at the first nonhelical positions in both termini of the helices, namely, N-cap and C-cap. The statistical analysis of the propensity values has shown that pairing is significantly dependent on the type of the amino acids and on the position of the pair. A few sequences of three and four amino acids were selected and their high prevalence in helices is outlined in this work. The Glu-Lys-Tyr-Pro sequence shows a peculiar distribution in proteins, which may suggest a relevant structural role in alpha-helices when Pro is located at the C-cap position. A bioinformatics tool was developed, which updates automatically and periodically the results and makes them available in a web site.

Alpha-helix stabilization by alanine relative to glycine: roles of polar and apolar solvent exposures and of backbone entropy

The energetics of alpha-helix formation are fairly well understood and the helix content of a given amino acid sequence can be calculated with reasonable accuracy from helix-coil transition theories that assign to the different residues specific effects on helix stability. In internal helical positions, alanine is regarded as the most stabilizing residue, whereas glycine, after proline, is the more destabilizing. The difference in stabilization afforded by alanine and glycine has been explained by invoking various physical reasons, including the hydrophobic effect and the entropy of folding. Herein, the contribution of these two effects and that of hydrophilic area burial is evaluated by analyzing Ala and Gly mutants implemented in three helices of apoflavodoxin. These data, combined with available data for similar mutations in other proteins (22 Ala/Gly mutations in alpha-helices have been considered), allow estimation of the difference in backbone entropy between alanine and glycine and evaluation of its contribution and that of apolar and polar area burial to the helical stabilization typically associated to Gly-->Ala substitutions. Alanine consistently stabilizes the helical conformation relative to glycine because it buries more apolar area upon folding and because its backbone entropy is lower. However, the relative contribution of polar area burial (which is shown to be destabilizing) and of backbone entropy critically depends on the approximation used to model the structure of the denatured state. In this respect, the excised-peptide model of the unfolded state, proposed by Creamer and coworkers (1995), predicts a major contribution of polar area burial, which is in good agreement with recent quantitations of the relative enthalpic contribution of Ala and Gly residues to alpha-helix formation.

The interrelationships of side-chain and main-chain conformations in proteins

The accurate determination of a large number of protein structures by X-ray crystallography makes it possible to conduct a reliable statistical analysis of the distribution of the main-chain and side-chain conformational angles, how these are dependent on residue type, adjacent residue in the sequence, secondary structure, residue-residue interactions and location at the polypeptide chain termini. The interrelationship between the main-chain (phi, psi) and side-chain (chi 1) torsion angles leads to a classification of amino acid residues that simplify the folding alphabet considerably and can be a guide to the design of new proteins or mutational studies. Analyses of residues occurring with disallowed main-chain conformation or with multiple conformations shed some light on why some residues are less favoured in thermophiles.

Determination of alpha-helix N1 energies after addition of N1, N2, and N3 preferences to helix/coil theory

Surveys of protein crystal structures have revealed that amino acids show unique structural preferences for the N1, N2, and N3 positions in the first turn of the alpha-helix. We have therefore extended helix-coil theory to include statistical weights for these locations. The helix content of a peptide in this model is a function of N-cap, C-cap, N1, N2, N3, C1, and helix interior (N4 to C2) preferences. The partition function for the system is calculated using a matrix incorporating the weights of the fourth residue in a hexamer of amino acids and is implemented using a FORTRAN program. We have applied the model to calculate the N1 preferences of Gln, Val, Ile, Ala, Met, Pro, Leu, Thr, Gly, Ser, and Asn, using our previous data on helix contents of peptides Ac-XAKAAAAKAAGY-CONH2. We find that Ala has the highest preference for the N1 position. Asn is the most unfavorable, destabilizing a helix at N1 by at least 1.4 kcal mol(-1) compared to Ala. The remaining amino acids all have similar preferences, 0.5 kcal mol(-1) less than Ala. Gln, Asn, and Ser, therefore, do not stabilize the helix when at N1.

Side-chain structures in the first turn of the alpha-helix

The first three residues at the N terminus of the alpha-helix are called N1, N2 and N3. We surveyed 2102 alpha-helix N termini in 298 high-resolution, non-homologous protein crystal structures for N1, N2 and N3 amino acid and side-chain rotamer propensities and hydrogen-bonding patterns. We find strong structural preferences that are unique to these sites. The rotamer distributions as a function of amino acid identity and position in the helix are often explained in terms of hydrogen-bonding interactions to the free N1, N2 and N3 backbone NH groups. Notably, the "good N2" amino acid residues Gln, Glu, Asp, Asn, Ser, Thr and His preferentially form i, i or i,i+1 hydrogen bonds to the backbone, though this is hindered by good N-caps (Asp, Asn, Ser, Thr and Cys) that compete for these hydrogen bond donors. We find a number of specific side-chain to side-chain interactions between N1 and N2 or between the N-cap and N2 or N3, such as Arg(N-cap) to Asp(N2). The strong energetic and structural preferences found for N1, N2 and N3, which differ greatly from positions within helix interiors, suggest that these sites should be treated explicitly in any consideration of helical structure in peptides or proteins.

A helix propensity scale based on experimental studies of peptides and proteins

The average globular protein contains 30% alpha-helix, the most common type of secondary structure. Some amino acids occur more frequently in alpha-helices than others; this tendency is known as helix propensity. Here we derive a helix propensity scale for solvent-exposed residues in the middle positions of alpha-helices. The scale is based on measurements of helix propensity in 11 systems, including both proteins and peptides. Alanine has the highest helix propensity, and, excluding proline, glycine has the lowest, approximately 1 kcal/mol less favorable than alanine. Based on our analysis, the helix propensities of the amino acids are as follows (kcal/mol): Ala = 0, Leu = 0.21, Arg = 0.21, Met = 0.24, Lys = 0.26, Gln = 0.39, Glu = 0.40, Ile = 0.41, Trp = 0.49, Ser = 0.50, Tyr = 0. 53, Phe = 0.54, Val = 0.61, His = 0.61, Asn = 0.65, Thr = 0.66, Cys = 0.68, Asp = 0.69, and Gly = 1.

Intrahelical side chain interactions in alpha-helices: poor correlation between energetics and frequency

Polypeptide sequences in proteins may increase their tendency to adopt helical conformations in several ways. One is the recruiting of amino acid residues with high helical propensity. Another is the appropriate distribution of residues along the helix to establish stabilising side chain interactions. The first strategy is known to be followed by natural proteins because amino acids with high helical propensity are more frequent in alpha-helices. If proteins also followed the second strategy, stabilising amino acid pairs should be more frequent than others. To test this possibility we compared empirical energies of side chain interactions in alpha-helices with statistical energies calculated from a data base of proteins with low homology. We find some correlation between the stability afforded by the pairs and their relative abundance in alpha-helices but the realisation of energetic preferences into statistical preferences is very low. This indicates that natural alpha-helices do not regularly use intrahelical side chain interactions to increase their stability.

Phosphorylation destabilizes alpha-helices

Phosphorylation of threonine destabilizes the leucine zipper of a bZIP protein by 4.6 kcal mol-1 dimer-1, which reduces DNA binding 100-fold. This decrease in stability reflects the low alpha-helix forming propensity of a phosphorylated threonine.

Prediction of the secondary structure content of globular proteins based on structural classes

The prediction of the secondary structure content (alpha-helix and beta-strand content) of a globular protein may play an important complementary role in the prediction of the protein's structure. We propose a new prediction algorithm based on Chou's database [Chou (1995), Proteins Struct. Funct. Genet. 21, 319]. The new algorithm is an improved multiple linear regression method, taking the nonlinear and coupling terms of the frequencies of different amino acids into account. The prediction is also based on the structural classes of proteins. A resubstitution examination for the algorithm shows that the average errors are 0.040 and 0.033 for the prediction of alpha-helix content and beta-strand content, respectively. The examination of cross-validation, the jackknife analysis, shows that the average errors are 0.051 and 0.044 for the prediction of alpha-helix content and beta-strand content, respectively. Both examinations indicate the self-consistency and the extrapolative effectiveness of the new algorithm. Compared with the other methods available currently, our method has the merits of simplicity and convenience for use, as well as a high prediction accuracy. By incorporating the prediction of the structural classes, the only input of our method is the amino acid composition of the protein to be predicted.

Prediction of protein folding class from amino acid composition

An empirical relation between the amino acid composition and three-dimensional folding pattern of several classes of proteins has been determined. Computer simulated neural networks have been used to assign proteins to one of the following classes based on their amino acid composition and size: (1) 4 alpha-helical bundles, (2) parallel (alpha/beta)8 barrels, (3) nucleotide binding fold, (4) immunoglobulin fold, or (5) none of these. Networks trained on the known crystal structures as well as sequences of closely related proteins are shown to correctly predict folding classes of proteins not represented in the training set with an average accuracy of 87%. Other folding motifs can easily be added to the prediction scheme once larger databases become available. Analysis of the neural network weights reveals that amino acids favoring prediction of a folding class are usually over represented in that class and amino acids with unfavorable weights are underrepresented in composition. The neural networks utilize combinations of these multiple small variations in amino acid composition in order to make a prediction. The favorably weighted amino acids in a given class also form the most intramolecular interactions with other residues in proteins of that class. A detailed examination of the contacts of these amino acids reveals some general patterns that may help stabilize each folding class.

A molecular thermodynamic approach to predict the secondary structure of homopolypeptides in aqueous systems

Under physiological conditions, many polypeptide chains spontaneously fold into discrete and tightly packed three-dimensional structures. The folded polypeptide chain conformation is believed to represent a minimum Gibbs energy of the system, governed by the weak interactions that operate between the amino acid residues and between the residues and the solvent. A semiempirical molecular thermodynamic model is proposed to represent the Gibbs energy of folding of aqueous homopolypeptide systems. The model takes into consideration both the entropy contribution and the enthalpy contribution of folding homopolypeptide chains in aqueous solutions. The entropy contribution is derived from the Flory-Huggins expression for the entropy of mixing. It accounts for the entropy loss in folding a random-coiled polypeptide chain into a specific polypeptide conformation. The enthalpy contribution is derived from a molecular segment-based Non-Random Two Liquid (NRTL) local composition model [H. Renon and J. M. Prausnitz (1968) AIChE J., Vol. 14, pp. 135-142; C.-C. Chen and L. B. Evans (1986) AIChE J., Vol. 32, pp. 444-454], which takes into consideration of the residue-residue, residue-solvent, and solvent-solvent binary physical interactions along with the local compositions of amino acid residues in aqueous homopolypeptides. The UNIFAC group contribution method [A. Fredenslund, R. L. Jones, and J. M. Prausnitz (1975) AIChE J., 21, 1086-1099; A. Fredenslund, J. Gmehling, and P. Rasmussen (1977) Vapor-Liquid Equilibrium Using UNIFAC, Elsevier Scientific Publishing Company, Amsterdam], developed originally to estimate the excess Gibbs energy of solutions of small molecules, was used to estimate the NRTL binary interaction parameters. The model yields a hydrophobicity scale for the 20 amino acid side chains, which compares favorably with established scales [Y. Nozaki and C. Tanford (1971) Journal of Biological Chemistry, Vol. 46, pp. 2211-2217; E. B. Leodidis and T. A. Hatton (1990) Journal of Physical Chemistry, Vol. 94, pp. 6411-6420]. In addition, the model generates qualitatively correct thermodynamic constants and it accurately predicts thermodynamically favorable folding of a number of aqueous homopolypeptides from random-coiled states into alpha-helices. The model further facilitates estimation of the Zimm-Bragg helix growth parameter s and the nucleation parameter sigma for amino acid residues [B. H. Zimm and J. K. Bragg (1959) Journal of Chemical Physics, Vol. 31, pp. 526-535]. The calculated values of the two parameters fall into the ranges suggested by Zimm and Bragg.

Predicting protein secondary structure content. A tandem neural network approach

A priori knowledge of secondary structure content can be of great use in theoretical and experimental determination of protein structure. We present a method that uses two computer-simulated neural networks placed in "tandem" to predict the secondary structure content of water-soluble, globular proteins. The first of the two networks, NET1, predicts a protein's helix and strand content given information about the protein's amino acid composition, molecular weight and heme presence. Because NET1 contained more adjustable parameters (network weights) than learning examples, this network experienced problems with memorization, which is the inability to generalize onto new, never-seen-before examples. To overcome this problem, we designed a second network, NET2, which learned to determine when NET1 was in a state of generalization. Together, these two networks produce prediction errors as low as 5.0% and 5.6% for helix and strand content, respectively, on a set of protein crystal structures bearing little homology to those used in network training. A comparison between three other methods including a multiple linear regression analysis, a non-hidden-node network analysis and a secondary structure assignment analysis reveals that our tandem neural network scheme is, indeed, the best method for predicting secondary structure content. The results of our analysis suggest that the knowledge of sequence information is not necessary for highly accurate predictions of protein secondary structure content.

Secondary structure prediction and determination of proteins--a review

The rapid increase in sequence data in combination with a greater understanding of the forces regulating protein structure has been the impetus for an upsurge in the development of theoretical prediction methods. These methods have afforded protein chemists the ability to identify and quantify the various secondary structures along the protein chain. Concurrently, various physico-chemical techniques have been developed such as nuclear Overhauser enhancement n.m.r. and laser Raman spectroscopy. In addition, traditional methods such as infrared and circular dichroism spectroscopy have been refined. Although both predictive and physico-chemical techniques are limited in the types of secondary structure they are capable of determining, they have provided valuable information with regards to protein folding and topology in the absence of X-ray data, and have formed the basis for the development of improved methods for secondary structure determination. This paper reviews some of the predictive and physico-chemical methods presently used to determine protein secondary structure.

A program for prediction of protein secondary structure from nucleotide sequence data: application to histocompatibility antigens

A computer program is described which, given a nucleotide or an amino acid sequence, outputs protein secondary structure prediction curves as well as hydrophobicity and charged-residue profiles. The program allows for cumulative averaging of properties (secondary structure propensities, hydrophobicity and charge profiles) from several homologous primary structures, a novel concept shown to improve the predictive accuracy. The use of the program is demonstrated on a set of nucleotide and amino acid sequences from human and murine histocompatibility antigens of class I and II. The last extracellular domains of both class I and II antigens (alpha 3 of class I, alpha 2 and beta 2 of class II) and the beta 2-microglobulin domain are predicted to consist of seven anti-parallel beta-strands, in accord with previous claims of homology between these domains and the constant domains of immunoglobulin chains. The remaining extracellular domains are all proposed to form an anti-parallel, four-stranded beta-sheet with one of its faces being covered by alpha-helices and/or structureless segments ("open face sandwiches").

Recent studies on the structure of human serum low-and high-density lipoproteins

Several methods have been recently developed for the preparation of soluble apo-low-density lipoprotein (apoLLL) Delipidation of LDL alters the immunochemical activity of the molecule and causes an apparent increase in random structure. Despite these changes, soluble apoLDL retains prominent immunological and optical characteristics of native LDL. The lipid of LDL appears to stabilize the protein conformation of LDL derivatives. Optical measurements suggest that native LDL contains a significant amount of pleatedsheet, antiparallel chain beta-structure in addition to random and probably some helical structure, while HDL, by optical criteria, is relatively richer in the alpha-helical conformation. Two proteins containing C-terminal glutamine and C-terminal threonine have been isolated from high-density lipoprotein (HDL). Circular dichroic measurements and total amino acid content are consistent with a greater helical content of apoHDL-Thr than apoHDL-Gln. The techniques of nuclear magnetic resonance and electron spin resonance have been used to probe the structure of LDL and HDL. The presence of protein does not seem to exert a constraining effect on the proton resonance of lipoprotein lipids, while the presence of lipid does appreciably constrain nitroxide tumbling in spin-labeled lipoprotein protein, the constraint being relatively greater in HDL than in LDL. Existing structural evidence is consistent with lipoprotein models in which protein and phospholipid occupy the surface and other lipids are more internal.