Author: Shilpa Shiragannavar
Protein secondary structure prediction refers to the prediction of the conformational state of each amino acid residue of a protein sequence as one of the three possible states, namely, helices, strands, or coils, denoted as H, E, and C, respectively. The prediction is based on the fact that secondary structures have a regular arrangement of amino acids, stabilized by hydrogen bonding patterns.
Predicting protein secondary structures has a number of applications.
- It can be useful for the classification of proteins and for the separation of protein domains and functional motifs.
- Secondary structures are much more conserved than sequences during evolution. As a result, correctly identifying secondary structure elements (SSE) can help to guide sequence alignment or improve existing sequence alignment of distantly related sequences.
- Secondary structure prediction is an intermediate step in tertiary structure prediction as in threading analysis.
SECONDARY STRUCTURE PREDICTION FOR GLOBULAR PROTEINS
The secondary structure prediction methods can be either ab initio based, which make use of single sequence information only, or homology based, which make use of multiple sequence alignment information. The ab initio methods, which belong to early generation methods, predict secondary structures based on statistical calculations of the residues of a single query sequence. The homology-based methods do not rely on statistics of residues of a single sequence, but on common secondary structural patterns conserved among multiple homologous sequences.
Ab Initio–Based Methods
This type of method predicts the secondary structure based on a single query sequence. It measures the relative propensity of each amino acid belonging to a certain secondary structure element. The propensity scores are derived from known crystal structures. Examples of ab initio prediction are the Chou–Fasman and Garnier, Osguthorpe, Robson (GOR) methods. The ab initio methods were developed in the 1970s when protein structural data were very limited. The statistics derived from the limited data sets can therefore be rather inaccurate. The methods are simple and are often used to illustrate the basics of secondary structure prediction.
The Chou–Fasman algorithm ( http://fasta.bioch.virginia.edu/fasta/chofas.htm ) determines the propensity or intrinsic tendency of each residue to be in the helix, strand, and β-turn conformation using observed frequencies found in protein crystal structures (conformational values for coils are not considered). For example, it is known that alanine, glutamic acid, and methionine are commonly found in α-helices, whereas glycine and proline are much less likely to be found in such structures.
The calculation of residue propensity scores is simple. Suppose there are n residues in all known protein structures from which m residues are helical residues. The total number of alanine residues is y of which x are in helices. The propensity for alanine to be in helix is the ratio of the proportion of alanine in helices over the proportion of alanine in overall residue population (using the formula[ x/m]/[y/n]). If the propensity for the residue equals 1.0 for helices (P[α-helix]), it means that the residue has an equal chance of being found in helices or elsewhere. If the propensity ratio is less than 1, it indicates that the residue has less chance of being found in helices. If the propensity is larger than 1, the residue is more favoured by helices. Based on this concept, Chou and Fasman developed a scoring table(Secondary structure element score) listing relative propensities of each amino acid to be in an α-helix, a β-strand, or a β-turn.
Prediction with the Chou–Fasman method works by scanning through a sequence with a certain window size to find regions with a stretch of contiguous residues each having a secondary structure elements score to make a prediction. For α-helices, the window size is six residues, if a region has four contiguous residues each having P(α-helix) > 1.0, it is predicted as an α-helix. The helical region is extended in both directions until the P(α-helix) score becomes smaller than 1.0. That defines the boundaries of the helix. For β-strands, scanning is done with a window size of five residues to search for a stretch of at least three favored β-strand residues. If both types of secondary structure predictions overlap in a certain region, a prediction is made based on the following criterion: if _P(α) > _P(β), it is declared as an α-helix; otherwise, a β-strand.
The GOR method (http://fasta.bioch.virginia.edu/fasta www/garnier.htm) is also based on the “propensity” of each residue to be in one of the four conformational states, helix (H), strand (E), turn (T), and coil (C). Instead of using the propensity value from a single residue to predict a conformational state; it takes short-range interactions of neighbouring residues into account. It examines a window of every seventeen residues and sums up propensity scores for all residues for each of the four states resulting in four summed values. The highest scored state defines the conformational state for the centre residue in the window (ninth position).The GOR method has been shown to be more accurate than Chou–Fasman because it takes the neighbouring effect of residues into consideration.
This type of method combines the ab initio secondary structure prediction of individual sequences and alignment information from multiple homologous sequences (>35% identity). The idea behind this approach is that close protein homologs should adopt the same secondary and tertiary structure. When each individual sequence is predicted for secondary structure using a method similar to the GOR method, errors and variations may occur. Evolutionary conservation indicates that there should be no major variations for their secondary structure elements. Therefore, by aligning multiple sequences, information of positional conservation is revealed. Because residues in the same aligned position are assumed to have the same secondary structure, any inconsistencies or errors in prediction of individual sequences can be corrected using a majority rule. This homology based method has helped improve the prediction accuracy by another 10% over the second-generation methods.
Prediction with Neural Networks
A neural network is a machine learning process that requires a structure of multiple layers of interconnected variables or nodes. In secondary structure prediction, the input is an amino acid sequence and the output is the probability of a residue to adopt a particular structure. Between input and output are many connected hidden layers where the machine learning takes place to adjust the mathematical weights of internal connections. The neural network has to be first trained by sequences with known structures so it can recognize the amino acid patterns and their relationships with known structures. During this process, the weight functions in hidden layers are optimized so they can relate input to output correctly.
When multiple sequence alignments and neural networks are combined, the result is further improved accuracy. The following lists several frequently used for prediction algorithms available as web servers.
PHD (Profile network from Heidelberg ; http://dodo.bioc.columbia.edu/predict protein/submit def.html) is a web-based program that combines neural network with multiple sequence alignment. It first performs a BLASTP of the query sequence against a non redundant protein sequence database to find a set of homologous sequences, which are aligned with the MAXHOM program (a weighted dynamic programming algorithm performing global alignment). The resulting alignment in the form of a profile is fed into a neural network that contains three hidden layers. The first hidden layer makes raw prediction based on the multiple sequence alignment by sliding a window of thirteen positions. As in GOR, the prediction is made for the residue in the center of the window. The second layer refines the raw prediction by sliding a window of seventeen positions, which takes into account more flanking positions. This step makes adjustments and corrections of unfeasible predictions from the previous step. The third hidden layer is called the jury network, and contains networks trained in various ways. It makes final filtering by deleting extremely short helices (one or two residues long) and converting them into coils. After the correction, the highest scored state defines the conformational state of the residue.
About Author / Additional Info: