1 HCAM - Hidropathy Clustering Assisted Method

Prediction of the native structure of newly characterised proteins is a central problem in molecular biology. This is why development of theoretical methods for identifying folding of polypeptide chains has become an important field of research over the past few years, permitting prediction algorithms based on many new approaches to be developed. Water soluble globular proteins are known to fold into structures in which hydrophobic residues are disposed on the inside while lateral chains of hydrophilic residues are exposed to the solvent. In native conformation, apolar residues tend to be disposed in the core, while polar residues tend to occupy the surface of the protein. Moreover, these proteins typically have an abundance of secondary structures such as α-helices and β-sheets connected by flexible loops that often play a role in determining biological diversity of function and activity. Despite the many interactions occurring between distant parts of the amino acid chain, a large variety of different protein sequences seem to fold into similar structural motifs and each structural element has distinct fields of preference for different amino acids, suggesting the presence of a finite number of folds, even thought the astronomical variety of possible combinations of residues composing polypeptide chains. In the course of evolution, the three-dimensional structure of certain folds seems conserved to a greater extent than the corresponding amino acid sequence, since substitution of certain residues that tend to stabilise the folding motifs may be compensated by other substitutions that give the structure stability. Moreover, several recurrent structural motifs retain specific biological functions. Since our capacity to predict the three-dimensional structure of a protein is limited by lack of a general paradigm, most prediction methods have focused on identification of both secondary and tertiary structure of proteins by computational approaches. Although several computational methods made a great effort to predict the whole proteins structure at high atomic resolution, template based search by sequence similarity (comparative modelling) and conformational search (ab initio modelling) are the strategies so far mostly used to obtain the tertiary structure of amino acid sequences and are mainly directed at predicting certain conserved structural sub-domains, as well as fragment-search based methods that focused on the prediction of well known super-secondary folds. Several properties shared by certain protein structural motifs could suggest strategies for the prediction of large populations of recurrent folds. For example, a case study based on exhaustive analysis of crystallized protein structures indicates that the main building blocks of globular proteins actually consist of closed loops of standard length. Protein structure can therefore be seen as a compact array of closed loops that can be grouped into two main sub-populations with 11-15 and 25-35 amino acids, respectively, that follow each other in strict linear order (see Figure 1). The presence of certain sequence motifs typical of closed loops enabled certain “prototypes” common to proteins of many bacteria to be identified, suggesting the existence of a proteomic code. The extremities of closed loops are often in close contact with each other (distance C-C < 10 Å) and are characterised by 3-5 hydrophobic and/or non polar residues, often corresponding to “hydrophobic folding units” (HFUs) of globular proteins.

Typical closed loops of 25-35 amino acids exhibit similar features for small and large proteins. The selective pressure common to closed loops is not only reflected in structure but also in the typical distance between hydrophobic residues, which is equal to the distance between the two extremities of the folding motif. Prediction of these extremities can therefore be obtained by means of physicochemical parameters such as the Kyte-Doolittle hydropathy scale, which is particularly indicated for revealing certain properties of globular proteins. Hydropathy scales are currently used to inspecting the hydrophobic character of proteins in order to reveal families of transmembrane helices, potential antigenic sites and regions that could be exposed on the protein surface. Moreover, the prediction power of hydropathy profile lies in the possibility of clarifying evolutionary relationships more distant than those obtained by comparison of amino acid sequences, but it has not yet been narrowed to the detection of secondary structure profiles shared by recurrent super-secondary structure motifs of both globular and transmembrane proteins. Here we endeavour to combine several analyses for obtaining fingerprints for identification of position and length of secondary structures placed in protein elementary building blocks, including closed loops and other typical sub-domains of polypeptide chains. Thus, we introduced the notion of Hydrophobic Assistance For Secondary Structure Detection and we decided to define our algorithm as “Hydropathy Clustering Assisted Method” (HCAM) because of the main contribution given by hydropathy analysis. HCAM results can be read as a whole protein prediction map, but the association of secondary structures with previously collected hydropathy patterns should lead toward targeted super-secondary motifs identification. In secondary structure prediction, any ambiguities caused by overlap of hydropathy patterns of certain types can be resolved by complementary statistical analysis based on confidence levels derived from the frequencies of particular amino acid motifs and by use of hydrophilic and hydrophobic residue patterns (binary patterns) that play an important role in protein architecture and can be divided in two classes in relation to whether they preferentially form α-helices or β-sheets.

The sequences of each class of protein building motifs were processed by multiple alignment analysis using the Gonnet substitution matrix in the programme ClustalW (see Figures 2a and 2b). Secondarily, the secondary structure members (α-helices, β-strands and coils) of each class were partitioned in clusters containing amino acid sequence motifs and the corresponding hydropathy values. Loop sequences linking motifs such as βαβ barrels, α-hairpins, αβ-barrels, etc. were partitioned.

1.2 Hydropathy profile analysis

The Kyte-Doolittle scale combined with a 5-amino acid sliding window is a useful instrument for revealing the topological disposition of secondary structures inside several building blocks. For example, careful observation of the hydropathy profile of prokaryotic globular proteins showed many β-barrels with two lateral hydrophobic “icebergs” (positive hydrophobic domains) 3-6 amino acids long, flanking a central negative area of variable length, generally not less than 3 amino acids; the two icebergs often correspond roughly to residues forming β-strands of β-barrels, whereas the negative central domain is often associated with hairpin folds. Visual analysis of βαβ barrels also revealed hydrophobic icebergs associated with β-strands and a ragged central domain, associated with α-helices, having positive peaks separated by minima, sometimes less than zero (see Figure 3a-b). Small continuous hydrophobic domains may also separate icebergs from the ragged central domain; these domains often correspond to loops connecting β-strands and α-helices. Similarly, observation of the profile of α-hairpin motifs often revealed two ragged domains of peaks (α-helices) surrounding a domain with negative hydropathy (hairpin). However, certain domains corresponding to α-helices may sometimes be mistaken for hydrophobic icebergs erroneously associated with β-strands, due to a contiguous series of hydrophobic residues, especially Val, Ala and Leu, within the structural motif (see Figures 4a-d). Hydrophobic portions corresponding to coils can also sometimes have positive values (presence of Gly and Ala), but hydropathic peak form is generally different from that of icebergs of β-strands. Thus, profile pattern is not unambiguous, but nevertheless provides more or less detailed information on the position of certain structural motifs.

1.3 Statistical analysis of amino acid sequences

To compensate for the ambiguities encountered in hydropathy profiles, a complementary analysis of secondary structures, based on amino acid frequencies in sequence clusters, was conducted. We chose to calculate the frequency of 20 x 20 combinations of dipeptides and the frequency of a large amount of tripeptides observed in structural clusters formed by α-helices, β-strands, coils and β-hairpins. Permutations of a sequence of n amino acids were considered. If aa₁, aa₂, aa₃, ...,aa_i,..., aa_n are the amino acids of the sequence, the linear permutations of the n-1 dipeptides will be: aa₁ aa₂, aa₂ aa₃, ..., aa_i-1 aa_i, ..., aa_n-1 aa_n, and those of the n-2 tripeptides will be: aa₁ aa₂ aa₃, ..., aa_i-1 aa_i aa_i+1, ..., aa_n-2 aa_n-1 aa_n, where n ≥ 3.

1.4 Statistical analysis of binary patterns

Protein sequences can be represented as binary patterns of polar (p) and non polar (n) amino acids. The linear sequence of amino acids can therefore have particular dispositions of polar and non polar residues, which may promote formation of particular secondary structures. These patterns have been used as a binary code for protein design, where the precise identity of residue composing the binary sequence can vary with certain flexibility. We therefore used binary patterns to define guidelines for identifying secondary structures in amino acid sequences, calculating the frequency of the 23 combinations of binary motifs obtained from the linear permutations of tripeptides. The residues included in the category of polar amino acids were: Arg, Lys, Asp, Glu, Asn, Gln and His, whereas non polar amino acids were: Phe, Leu, Ile, Met, Val, Trp and Cys.
Analysis on an independent set of 150 globular proteins showed that the prediction method was particularly sensitive for each class of secondary structure and demonstrated that HCAM can also be used as a stand alone prediction program. Indeed, parameter Q3 (ratio of number of residues correctly predicted to total number of residues identified by the method) for elementary building blocks and closed loops was estimated at about 76% ± 5, while Q3 for the entire protein sequences was estimate at about 73%± 5. Moreover, the predictive estimate of secondary structure of building blocks motifs can considerably improve with manual intervention by expert operators, based on multiple alignment consensus methods and visual analysis of peaks of the overall hydropathy profile of the sequences: in this case Q3 may be as high as 81% since the help provided by the method makes manual identification of hydrophobic domains simpler. Although hydrophobic domains cause interpretative incongruence if viewed in the global topological context of the polypeptide sequence, they may suggest the exact correspondence with a secondary structural motif.