Method and System for Protein Modeling

This invention relates generally to a computer-implemented method and system for modeling the three-dimensional structure of a protein, and, more specifically, to protein modeling using the three-dimensional structure of a template protein and sequence alignment between the protein to be modeled and the template protein.

The function of a protein is related to its three-dimensional structure. The three-dimensional structure of some proteins can be determined using X-ray crystallography or Nuclear Magnetic Resonance (NMR). However, for other proteins, these methods cannot be used to determine the three-dimensional structure. When these methods cannot be used, computer-based protein modeling techniques have been used with some success. These protein modeling techniques use the known three-dimensional structure of a homologous protein to approximate the structure of another protein.

In one such technique, the known three-dimensional structures of the proteins in a given family are superimposed to define the structurally conserved regions in that family. Among the members of a given family, there is considerable variation in the conformations of regions located between two consecutive structurally conserved regions, and thus, these regions are called the variable regions. These variable regions essentially contribute to the identity of a protein in its family. However, the modeling of the variable regions has been unsatisfactory.

Conventional homology modeling techniques have been used routinely to build models of proteases and antibodies. These techniques generally model the three-dimensional position of amino acids in structurally conserved regions by taking the Cartesian coordinates from homologous amino acids in a template protein with known three-dimensional structure. For the amino acids in the variable regions, these techniques take suitable loops from the Protein Data Bank (PDB). (The PDB is a collection of protein information relating to the known structure of protein. The PDB is administered by Brookhaven National Laboratory.) Although these techniques are generally successful in modeling the structurally conserved regions for structurally undefined members of the family, these techniques have been unsuccessful in modeling of the variable regions. Since variable regions and structurally conserved regions of the model protein come from different protein structures, high energy short contacts are often found in the models. These high energy contacts are usually between intervariable regions that are grafted from different known protein structures. In cases where the sequence identity in the structurally conserved regions between the template and the model protein is weak, the interior amino acids are also susceptible to short contacts. Generally, these short contacts are removed by performing rotation around single bonds using interactive graphics, which is a tedious, and at times, an impractical procedure. Energy minimization has been used to relax strains in a model. However, the minimization procedure leads to structures that are trapped in local minima and relies entirely on the integrity of the starting structure.

Certain proteins with very weak sequence identities fold into similar three-dimensional structures. For example, the three-dimensional structures of a number of helical cytokines fold in similar three-dimensional topology in spite of weak sequence homology. Members of helical cytokine family not only show diversity in their disulfide topology but also the disulfide crosslinks are part of the variable regions. Besides the helical cytokines, there are other protein families with weak sequence identities and non-homologous disulfides in the variable regions. The prior homology modeling techniques produce unsatisfactory modeling of these families because of the absence of sequence homology in the structurally conserved regions and the presence of non-homologous disulfide crosslinks in the variable regions.

It is an object of the present invention to provide a method and system for modeling the three-dimensional structure of members of a protein family/superfamily from the known three-dimensional structure of one or more members of that family.

It is another object of the present invention to provide a method and system for modeling the three-dimensional structure of a protein by using position information of all topologically equivalent atoms in a template protein.

It is another object of the present invention to provide a method and system for generating a sequence alignment between two proteins based on the resulting three-dimensional structure model.

It is another object of the present invention to provide a method and system for generating a protein model that minimizes short contacts.

It is another object of the present invention to provide a method and system for modeling proteins with weak sequence identity with a template protein.

It is another object of the present invention to provide a method and system for modeling variable regions between two structurally conserved regions of a protein.

It is another object of the present invention to provide a method and system for modeling variable regions constrained by disulfide crosslinks of a protein.

These and other objects, which will become apparent as the invention is more fully described below, are provided by a computer-implemented method and system for modeling a three-dimensional structure of a model protein. In a preferred embodiment, the modeling is based upon a three-dimensional structure of a template protein and an amino acid sequence alignment of the model protein and the template protein. The proteins comprise a plurality of amino acids having backbone atoms and side chain atoms. For each amino acid in the model protein, when the template protein has an amino acid aligned with the amino acid of the model protein, the position of each backbone atom of the amino acid of the model protein is established based on the position of a topologically equivalent backbone atom in the aligned amino acid of the template protein. Then, the inter-atomic distance constraints for each pair of atoms with an established position is generated. Finally, the position of each atom in the model protein is set so that the inter-atomic distances are in accordance with the constraints.

FIG. 1 is an overall block diagram of the input and output for the protein modeler.

FIG. 2 is an overall flow diagram of the protein modeler.

FIG. 3 is a flow diagram of a routine to generate a TEA position data structure for the model protein.

FIG. 4 is a routine to generate distance constraints for topologically equivalent atoms.

FIG. 5 is a routine to generate standard constraints for non-topologically atoms.

FIG. 6 is a flow diagram showing the DGEOM process.

The present invention provides a computer-based method and system for modeling the three-dimensional structure of a protein (model protein). In a preferred embodiment, the protein modeler generates the three-dimensional structure of the model protein based on a template protein with a known three-dimensional structure and an amino acid sequence alignment between the model protein and the template protein. Using the template protein and the sequence alignment, the protein modeler generates a variety of inter-atom distance constraints for conserved regions and standard and chemical constraints for variable regions, and allows for the input of miscellaneous constraints (e.g., disulfide crosslinks). The protein modeler then generates a three-dimensional structure for the model protein using known techniques of distance geometry to ensure compliance with the constraints. Appendix A contains a further description of the present invention.

The protein modeler generates the inter-atomic distance constraints for the model protein based on the position of topologically equivalent atoms (TEA) in the template protein. The protein modeler determines whether each atom in the model protein is topologically equivalent to an atom in the template protein. An atom of an amino acid of the model protein is topologically equivalent when the amino acid of the model protein aligns with an amino acid of the template protein and the amino acid of the template protein has a corresponding atom. If the aligned amino acids are identical, then each atom of the model amino acid corresponds to the atom of the same name in the template amino acid. For example, the C-alpha atom of the model amino acid corresponds to the C-alpha atom of the template amino acid, and the C-beta atom of the model amino acid corresponds to the C-beta atom of the template amino acid. If the aligned amino acids are not identical, then each atom of the model amino acid may or may not have a corresponding atom in the template amino acid. Each backbone atom (N, C-alpha, C', and O) has a corresponding atom. However, the correspondence between side chain atoms depends on the aligned amino acids. For example, if the model amino acid is arginine and the template amino acid is alanine, then the C-beta atoms correspond. However, the C-gamma, C-delta, and other atoms of the arginine side chain have no corresponding atoms in the alanine side chain.

The protein modeler sets the position of each topologically equivalent atom in the model protein to the same position as that of the topologically equivalent atom in the template protein. The protein modeler then calculates the distance between each pair of atoms that have topologically equivalent atoms. These distances are the inter-atomic distance constraints for the topologically equivalent atoms in the model protein. In one embodiment, these distance constraints can be relaxed by accounting for the overall homology between the model and template protein as discussed below.

The protein modeler generates standard constraints based on the planarity of a side chain, the C-alpha and C-beta chirality, and C-alpha inter-atomic distances.

In a preferred embodiment, the protein modeler assigns a helical conformation to each amino acid in the model protein and establishes a position for each atom. The protein modeler then uses well-known software packages to generate chemical constraints such as inter-atomic distances between 1-2 and 1-3 neighbors based on the connectivity of the model protein. The inter-atomic distance constraints are used to constrain the topology relating to non-topologically equivalent atoms (non-TEAs). The selection of a helical conformation is arbitrary and used to establish position information for the standard software package. One skilled in the art would appreciate that one could generate such 1-2 and 1-3 neighbor constraints without use of standard software packages, and, thus obviate the need to establish the position of each atom to generate these constraints.

Once the various constraints are established, the well-known distance geometry software, DGEOM (discussed below), is invoked to calculate the position of each atom in the model protein and thus generates the three-dimensional structure of the model protein.

FIG. 1 is an overall block diagram of the input and output for the protein modeler. The protein modeler 110 is preferably implemented on a computer system. In a preferred embodiment, the protein modeler 110 inputs template protein structure 111, the model protein partial structure 112, and the model/template protein sequence alignment. Based on the input information, the protein modeler generates a model protein structure 114. The format of the information defining the protein structures are in a format defined by the Protein Data Bank (PDB). Table 1 illustrates the PDB protein structure format.

Table 1 contains an atom portion and a connectivity portion. The atom portion specifies for each atom in the protein a sequential atom number (iAtom), the atom name (Atom), the containing amino acid name (AA), the containing amino acid sequence number (iAA), and the Cartesian coordinates (position) of the atom (X,Y,Z). The connectivity portion indicates the connectivity of the atoms in the protein. The connectivity specifies each of the (up to four) neighbors of an atom.

FIG. 2 is an overall flow diagram of the protein modeler. In step 201, the protein modeler generates an initial three-dimensional structure for the model protein. In a preferred embodiment, the protein modeler inputs model protein partial structure 112, which contains structure information in PDB format without position information. One skilled in the art would appreciate that the model protein partial structure 112 could be generated by the protein modeler based on the model protein amino acid sequence. The position information for each atom in the model protein is generated based on a helical conformation. This position information is used by DGEOM to determine chemical constraints between the atoms of the model protein. These chemical constraints are later used to constrain the position of non-topologically equivalent atoms. The arbitrary assignment of helical conformation to the peptide backbone and the choice of side chain rotomers does not impose any restriction on subsequent stages of the modeling procedure. In fact, non-helical conformations can also produce acceptable results. An acceptable helical conformation algorithm may be found Sundaram, K. and Srinivasan, S., Computer Programs Biomed, "Computer Simulated Modeling of Biomolecular Systems," Vol. 10, p. 29-34, 1979. As discussed above, this step is used to take advantage of predefined software packages that generate chemical constraints. However, the chemical constraints could be generated directly from the model protein amino acid sequence based on well-known chemical constraint principles such as 1-2 and 1-3 neighbor topology.

In step 202, the protein modeler generates a TEA position data structure for the model protein based on topologically equivalent atoms. The sequence alignment between the template and the model protein is used to generate the TEA position data structure for the protein model. The protein modeler retrieves the position information for each topologically equivalent atom in the template protein and stores the result in the TEA position data structure for the model protein. The protein modeler also indicates in the TEA position data structure which atoms have no topologically equivalent atom in the template protein. Table 2 illustrates a sample TEA position data structure (also known as a hybrid data structure because it contains information based on the model and template proteins). The TEA position data structure contains an entry for each atom in the model protein. Each entry contains the following fields. The derived field contains a `T` for a topologically equivalent atom and an `M` for a non-topologically equivalent atom. The fields model amino acid number (miAA), model amino acid name (mAA), model atom number (miAtom), and model atom name (mAtom) identify the atom of the model protein. The fields template amino acid number (tiAA), template amino acid name (tAA), template atom number (tiAtom), and template atom name (tAtom) identify the topologically equivalent atom in the template protein. The fields X, Y, and Z specify the position of a topologically equivalent atom.

In step 203, the protein modeler generates the distance constraints based on the TEA position data structure. The protein modeler determines the distance between each topologically equivalent atom. In step 204, the protein modeler generates standard constraints for non-topologically equivalent atoms. These standard constraints include planarity of side chains, 1-4 inter-C-alpha distances, and C-alpha and C-beta chirality. In step 205, the protein modeler inputs miscellaneous constraints. These miscellaneous constraints allow for the methods of the present invention to be expanded to include other constraints as they become known. For example, a miscellaneous constraint could be distance constraints imposed by disulfide crosslinks. In step 206, the protein modeler generates the three-dimensional structure for the model protein based on the constraints and the helical three-dimensional structure of the model protein. The protein modeler preferably employs the well-known program DGEOM to generate the three-dimensional structure for the model protein. The DGEOM program is one of various well-known programs that are based on the principles of distance geometry. Appendix B contains a description of DGEOM. A detailed explanation of the theory in practice of distance geometry can be found in Havel, T. F., Kuntz, I. D., and Crippen, G. M., Bull. Math Biol., "Theory and Practice of Distance Geometry," Vol. 45, pp. 666-720, 1983.

FIG. 3 is a flow diagram of a routine to generate a TEA position data structure for the model protein. This routine inputs the model and template sequence alignment and the three-dimensional structure of template protein, and outputs the TEA position data structure for the model protein. In step 301, the routine selects the next model and template amino acid from the sequence alignment. In step 302, if all the amino acids have already been selected, then the routine returns, else the routine continues at step 303. In step 303, if the model protein has an amino acid in the sequence position but the template protein has no amino acid, then the amino acid is part of a variable region and the routine continues at step 304, else the routine continues at step 305. In step 304, the routine adds an entry to the TEA position data structure for each atom of the model amino acid as a non-topologically equivalent atom and loops to step 301 to select the next amino acid in the sequence alignment. In step 305, if both the model and template proteins have an amino acid in the sequence alignment, then the routine continues at step 306 to determine which atoms between the amino acids are topologically equivalent, else the routine loops to step 301. In step 306, the routine selects the next atom in the model amino acid starting with a first atom. In step 307, if all the atoms in the model amino acid have already been selected, then the routine loops to step 301, else the routine continues at step 308. In step 308, if the template amino acid has a topologically equivalent atom to the selected atom of the model amino acid, then the routine continues at step 309, else the routine continues at step 310. In step 309, the routine adds an entry to the TEA position data structure for the selected atom as a topologically equivalent atom and loops to step 306. In step 310, the routine adds an entry to the TEA position data structure for the selected atom as a non-topologically equivalent atom and the routine loops to step 306.

Code Table 1 contains "C" programming language pseudocode corresponding to the flow diagram of FIG. 3. The data structure Align contains the protein sequence alignment. The data structure Helical contains data relating to the three-dimensional structure of a helical conformation of the model protein. However, the pseudocode does not use the helical position data. The data structure Template contains the data relating to the three-dimensional structure of the template protein. The data structure Hybrid contains the TEA position data structure. The fields names are the same as described above for the three-dimensional structure in PDB format and the TEA positional data structure. The pseudocode also uses the position identity matrix (PIM) defined in Table 2.

Table 2 is the position identity matrix for side chain atoms of amino acids. The rows of table 2 represent amino acids and the columns represent positional identity of the amino acids. Topologically equivalent atoms in different amino acids are aligned in columns with an entry of 1. For example, if alanine is replaced by arginine in the model, the position of the arginine atoms--N, C-alpha, C+, O, and C-beta--in the template protein are used. Although not shown, the position identity matrix also contains columns for each of the four backbone atoms.

FIG. 4 is a flow diagram of the routine to generate distance constraints for topologically equivalent atoms. The routine inputs the TEA position data structure of the model protein and outputs distance constraints for the topologically equivalent atoms. In step 401, the routine selects the next topologically equivalent atom in the TEA position data structure, starting with the first topologically equivalent atom. In step 402, if all the topologically equivalent atoms have already been selected, then the routine returns, else the routine continues at step 403. In steps 403 through 407, the routine loops determining the distance between the selected topologically equivalent atom and all other topologically equivalent atoms. In step 403, the routine chooses a next topologically equivalent atom after the selected topologically equivalent atom in the TEA position data structure. In step 404, if all the topologically equivalent atoms after the selected topologically equivalent atom have already been chosen, then the routine loops to step 401 to select another topologically equivalent atom, else the routine continues at step 405. In step 405, the routine calculates the distance between the chosen and selected topologically equivalent atoms. In step 406, the routine calculates the upper and lower bounds of the distance constraints. As discussed above, the upper and lower bounds can be relaxed based on the overall homology between the model and template proteins. In step 407, the routine outputs the distance constraints between the chosen and the selected topologically equivalent atoms and the routine loops to step 403. Code Table 2 contains "C" programming language pseudocode for the flow diagram of FIG. 4.

FIG. 5 is a flow diagram of a routine to generate standard constraints for non-topologically equivalent atoms. The routine inputs the TEA position data structure and outputs the standard constraints for non-topologically equivalent atoms. In step 501, the routine selects the next non-topologically equivalent atom in the TEA position data structure, starting with the first non-topologically equivalent atom. In step 502, if all the non-topologically equivalent atoms have already been selected, then the routine returns, else the routine continues at step 503. In step 503, if the selected non-topologically equivalent atom is part of a planar side chain, then the routine outputs a planar constraint in step 504 and continues at step 505. In step 505, if the selected non-topologically equivalent atom is a C-beta atom and the amino acid is isoleucine or threonine, then the routine outputs a C-beta chiral constraint in step 506 and continues at step 507. In step 507, if the selected non-topologically equivalent atom is a C-alpha atom and the amino acid is not glycine, then the routine outputs a C-alpha chiral constraint in step 508 and continues at step 509. In step 509, if the selected non-topologically equivalent atom is a C-alpha atom, then the routine outputs the distance constraints for the next C-alpha atom in the peptide chain. The routine then loops to step 501 to select the next non-topologically equivalent atom. Code Table 3 contains "C" programming language pseudocode for the flow diagram of FIG. 5.

FIG. 6 is a flow diagram of the utilization of the DGEOM software package. In step 601, DGEOM initializes an inter-atomic distance matrix based on the helical three-dimensional structure of the model protein. DGEOM determines inter-atomic distances based on chemical constraints such as inter-atomic distances between 1-2 and 1-3 neighbors. Although DGEOM establishes these distance constraints for all atoms within the model protein, the distance constraints for topologically equivalent atoms are eventually overridden by distance constraints imposed by the template protein. In step 602, DGEOM applies the distance constraints for the topologically equivalent atoms to the distance matrix. In step 603, DGEOM applies the standard constraints for non-topologically equivalent atoms to the distance matrix. In step 604, the routine applies miscellaneous constraints to the distance matrix. In step 605, DGEOM smoothes the upper and lower bounds in the distance matrix. In step 606, DGEOM generates the final three-dimensional structure of the model protein based on the smoothed distance matrix.

The following example illustrates the methods of the present invention. Table 3 contains a sequence alignment for the murine interleukin-4 and human interleukin-4 proteins. In this example, the three-dimensional structure of the human interleukin-4 protein is known, and the three-dimensional structure of the murine interleukin-4 (the model protein) protein is to be modeled. In this example, a template protein is derived from the human interleukin-4 protein. The variable regions in the template protein are preferably expanded to allow for flexibility in modeling the transitions between conserved and variable regions.

Continuing with the example, the protein modeler sets the position of each atom in the model protein with a topologically equivalent atom in the template protein to the position of the atom in the template protein. For example, the first aligned amino acids for these model and template proteins are both leucine. Since each atom in the model amino acid has a topologically equivalent atom in the template amino acid, the protein modeler sets the position of each atom. Once the position of each topologically equivalent atom in the model protein is established, then the protein modeler generates the interatomic distance constraints for each pair of atoms. The protein modeler then determines standard constraints for the non-topologically equivalent atoms. For example, since the first amino acid, histidine, in the model protein does not align with an amino acid in the template protein, each atom of histidine is a non-topologically equivalent atom. Thus, a planar constraint, a C-alpha chiral constraint, and an inter-C-alpha distance constraint are generated. The protein modeler then inputs miscellaneous constraints that are generated by a user of the protein modeler. These constraints may typically be based on disulfide crosslinks or other empirically determined constraints. The protein modeler also generates chemical constraints for the non-topologically equivalent atom. Finally, the protein modeler applies all the constraints using the principles of the distance geometry to arrive at a three-dimensional structure of the murine interleukin-4 protein.

Although the present invention has been described in terms of a preferred embodiment, it is not intended that the invention be limited to this embodiment. Modifications within the spirit of the invention will be apparent to those skilled in the art, the scope of the present invention is defined by the claims which follow.