JP3129202B2 - Sequence secondary structure prediction method and sequence secondary structure prediction device - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a method and apparatus for predicting a secondary structure of a sequence, and more particularly to predicting a secondary structure of a sequence from a nucleic acid sequence such as a protein or an RNA (ribonucleic acid) whose tertiary structure is unknown. The present invention relates to a method and apparatus for predicting a secondary structure of a sequence.

[0002]

2. Description of the Related Art Proteins include G (glycine), A (alanine), V (valine), I (isoleucine), L (leucine), F (phenylalanine), P (proline), M
(Methionine), S (serine), T (threonine), Y
(Tyrosine), W (tryptophan), N (asparagine), Q (glutamine), C (cysteine), D (aspartic acid), E (glutamic acid), K (lysine), R
(Arginine) and H (histidine) are polypeptides in which 20 types of residues are linearly linked by peptide bonds.

[0003] Linear polypeptides fold into complex tertiary structures. This is because the atoms of the main chain that are not continuous in the sequence form a weak bond such as a hydrogen bond or a disulfide bond (a dehydration bond between cysteine side chains). In the hydrogen bond, in the main chain, hydrogen bonded to oxygen and nitrogen is bonded to hydrogen and oxygen bonded to nitrogen in the other part of the sequence, respectively. Therefore, one residue is directly linked to one residue.

[0004] The structure determined by the pair includes a helical structure called α-helix and a linear structure called β-sheet. The α-helix is a stable structure formed by hydrogen bonding between the main chains of the i-th residue and the (i + 4) -th residue in a helix having a period of 3.5 to 6 residues. Beta sheet
The main chains are hydrogen-bonded to form a pair, and the pair continuously forms a film-like structure. There are two types of β sheets, a parallel β sheet and a parallel β sheet that are joined in the same direction. For analyzing which part contributes to the three-dimensional structure of a protein, see JP-A-5-28.
No. 2381 (Matsuo, "Protein molecular three-dimensional structure analyzer": hereinafter referred to as Document 1).

[0005] In the β sheet, one residue is apparently adjacent to two residues because the direction of hydrogen bonding is exactly opposite between residues adjacent to each other on the sequence. Further, at this time, by the molecules of the side chains of the bonded residues interfering with each other, a pair of residues that are likely to be paired and a pair of residues that are unlikely to be paired are determined. These are described in reference 2 (Branden, C., and Tooze, J., "Introductio
n to Protein Structure ", Garland Publishing, (199
1)), Japanese translation is "Introduction to Protein Structure", Kyoikusha (19)
92)).

Conventional systems for predicting β-sheets of proteins are roughly classified into two types. First, a statistical tendency in which residues are included in the β sheet is obtained in advance from a database,
Based on this, it is predicted whether several consecutive residues on the sequence are included in the β-sheet. These are described in Reference 3 (Cho
u, PY, Fasman, GD, "Prediction of protein confroma
tion ", Biochemistry, 13: 222-245, (1974).

[0007] As another method, the statistical tendency of forming a pair from a protein whose tertiary structure is known,
It is obtained in advance from a database or learned by a neural network, and it is predicted whether a β sheet is formed for all pairs on an array. For this, see Reference 4.
(Hubbard T., "Use of β-strand Interaction Pseudo-
Potentials in Protein Structure Prediction and Mod
elling ", Proceedingsof the 27th Annual Hawaii Inter
national Conference on System Sciences, pp. 336-334.
(1994)).

On the other hand, the RNA sequence has four types of bases, namely, adenine (A), uracil (U), and guanine (G).
And cytosine (C). These RNAs
Bases A and U, G and C, and unstable but G and U
Bind complementarily. By this binding, RNA has
A structure (stem structure) like a β-sheet of a protein appears. Differences from proteins are that there are no parallel sheets due to constraints between bases, and that each base can be adjacent to only one base at most (up to 2 for proteins).

[0009] Since RNA is simpler than protein, it is experimentally known about the binding energy between bases and the stabilizing energy depending on the stem length. Although a method of finding a shape having the lowest energy using dynamic programming is well known (Zuker method), a pseudo knot (Pseudo-
Knot) cannot be dealt with. These are described in reference 5 (Gribskov M. and Devereux J., "Se
quence Analysis Primer ", Stockton press, (1991)).

[0010]

However, none of the above-mentioned conventional secondary structure prediction methods determines whether or not it is possible to actually take a three-dimensional structure, and cannot be realized as a system output. There is a drawback that some are included. The analysis of how the interaction between individual local structures of a protein molecule contributes to the formation of a three-dimensional structure is described in Reference 1, but this document estimates the three-dimensional structure. There is no mention of what to do.

The present invention has been made in view of the above points,
A sequence secondary structure prediction method and a sequence secondary structure prediction device capable of plausible three-dimensional structure prediction by determining pairs forming pairs in a sequence while determining whether or not a three-dimensional structure can be actually taken. The purpose is to provide.

[0012]

In order to achieve the above-mentioned object, the method of the present invention uses a binding energy between constituent units of a protein and a nucleic acid sequence to obtain a protein or nucleic acid sequence whose input tertiary structure is unknown. After calculating the binding energies for all combinations of the two sub-arrays, the combination of the sub-arrays having the lowest sum of the binding energies is determined using a Hopfield neural network expressing the constraint of the three-dimensional structure, The input three-dimensional structure predicts the secondary structure of an unknown protein or nucleic acid sequence.

In the method of the present invention, the secondary structure of a protein or nucleic acid sequence whose tertiary structure is unknown is predicted in consideration of the restriction of tertiary structure. A more accurate sequence secondary structure can be predicted.

[0014] In addition to the above, the method of the present invention combines a simulation to determine whether the result obtained by the Hopfield neural network can be folded. Thereby, more accurate prediction of the sequence secondary structure can be performed.

In order to achieve the above object, the apparatus of the present invention comprises a binding energy holding unit for holding binding energy between constituent units of a protein and a nucleic acid sequence, and a protein or nucleic acid sequence whose input tertiary structure is unknown. And a binding energy for all combinations of the two partial sequences from the partial sequence combination generating unit using the binding energy from the binding energy holding unit. And a binding energy calculated by the binding energy calculation unit are set as initial values,
A configuration having a Hopfield neural network unit that predicts a combination of partial sequences that generate the lowest binding energy while satisfying the constraints of the three-dimensional structure as a sequence secondary structure of an unknown protein or nucleic acid sequence. It was done.

In the apparatus of the present invention, the Hopfield neural network unit determines the combination of the partial sequences that generate the lowest binding energy while satisfying the constraints of the three-dimensional structure, and the sequence of the protein or nucleic acid sequence whose unknown three-dimensional structure is unknown. Since the prediction is performed as a structure, a more accurate sequence secondary structure can be predicted as compared with a conventional prediction device that does not consider the restriction of the three-dimensional structure.

Further, the binding energy holding unit of the device of the present invention is characterized by β β expressed by a combination of partial sequences of proteins.
The sheets are classified into three types according to a parallel type, an anti-parallel type, and a hydrogen bond pattern, and hold the binding energy of each type. The binding energy calculation unit calculates the binding energy of each of the three types of binding energy. It is characterized by performing calculations.

Further, the Hopfield neural network section of the apparatus of the present invention is characterized in that β-sheets represented by combinations of partial sequences of proteins are classified into three types based on the pattern of parallel type, antiparallel type and hydrogen bond. The operation is performed using an energy function expressing a constraint for each type.

Furthermore, it is desirable for a more accurate prediction that the Hopfield neural network unit operates using an energy function that expresses the beginning and end of the β-sheet and expresses a constraint on the continuity of the β-sheet.

The Hopfield neural network unit operates using an energy function that expresses the beginning and end of an α-helix expressed by a combination of partial sequences of proteins, and expresses constraints on the continuity of the α-helix. However, it is preferable because the sequence secondary structure can be predicted in consideration of the restriction of the three-dimensional structure.

Further, the apparatus of the present invention further includes a folding simulation unit which receives a value of the operating Hopfield neural network unit as an input, performs a folding simulation, and feeds back the value of the Hopfield neural network unit. However, it is desirable in that the accuracy of prediction can be further improved.

[0022]

Next, embodiments of the present invention will be described with reference to the drawings.

FIG. 1 shows an overall configuration diagram of a first embodiment of an array secondary structure prediction device according to the present invention. As shown in the figure, the sequence secondary structure predicting apparatus includes a binding energy holding unit 11 for holding binding energy of each structural unit (amino acid in the case of protein, nucleic acid in the case of RNA), and a combination of partial sequences. Binding energy calculation unit 12 for calculating binding energy by means of a combination, subsequence combination generation unit 13 for generating all combinations of two subsequences of sequence 15 whose tertiary structure is unknown, and Hopfield neural network expressing constraints on steric structure And a network unit 14.

RN held by the binding energy holding unit 11
As shown in Reference 5, the binding energy of A has been obtained by a biochemical experiment. In addition, since the binding energy of a protein is not known experimentally, it is determined based on statistics of a protein whose tertiary structure is actually known. The method is known from document 4.

In the case of proteins, the partial tertiary structure 2
As the next structure, a helical structure called α-helix and β
There is a linear structure called a sheet. In these secondary structures, the binding energy between residues is estimated using statistics. For example, for the β sheet, the binding energy is determined for a pair of a 3-residue partial sequence and a 3-residue partial sequence. The three residues are tentatively determined from the number of databases of the three-dimensional structure, and partial sequences having other lengths can also be used.

The three-residue string-three-residue string pairs extending between strands in the β-sheet of the protein are classified into three types. The first type is an antiparallel β as shown in FIG.
In the 3-residue line-3 residue sequence pair in the sheet, there is a type H in which hydrogen bonds are present between the residues at both ends.
As shown in (b), 3 residue sequence in antiparallel β sheet-3
In the residue type pair, type A in which a hydrogen bond is present only between the central residues, and the third type is, as shown in FIG. Type P which is a pair
It is. The binding energy holding unit 11 in FIG.
It holds the binding energies of the two types.

In the example in which the three-residue sequence is used for the β sheet, the partial-sequence combination generating unit 13 shown in FIG. Generate all two combinations of The binding energy calculation unit 12 is a partial array combination generation unit 1
The respective binding energies for all combinations of the three residue sequences generated by 3 are calculated using the binding energies stored in the binding energy holding unit 11.

In the calculation of the binding energy, (N−
1) ^2/2 binding energy is computed. The reason for the doubling is that the three residue sequence is exchangeable.
If not exchangeable, do not multiply by 1/2 and (N-1) ²
Is calculated. In addition, the reason why (N-1) is used is that it is omitted in advance because the same three-residue sequence as the three-residue sequence never becomes a three-residue sequence-3 residue sequence pair. It is. In the case of proteins, since they are classified into the above three types, H, A and P, 3 × (N
-1) ^2/2 binding energy can be calculated.

The Hopfield neural network unit 14 of FIG. 1, (in the case of proteins, 3 × (N-1) 2/2 pieces of binding energy) is calculated by the binding energy calculation unit 12 the calculation result set as an initial value And operates using the same number of cells as the initial value. Also, 2 × N cells are used for the α helix and β sheet.

FIG. 3 shows a part of the Hopfield neural network unit 14, wherein FIG. 3A shows type H, FIG. 3B shows type A, and FIG. Hi _{, j} of the Hopfield neural network unit 14 shown in FIG. 3A is a three residue sequence (residue i-1, residue i, residue i + 1) and a three residue sequence (residue j-1). , Residue j,
Whether residue j + 1) has a type H bond, for example, [0.
0,1.0]. Hereinafter, the three residue sequence (residue i-1, residue i, residue i + 1) is the residue sequence i, and the three residue sequence (residue j-1, residue j, residue j + 1) is the residue sequence. Expressed as j.

Similarly, A _{i, j} shown in FIG. 3 (b) and P _{i, j} shown in FIG. 3 (c) indicate that the residue sequence i and the residue sequence j indicate the binding of type A and type P, respectively. It expresses whether you have.
That is, for example, “H _{i, j} ” shown by one circle in FIG.
A three-residue column pair of a residue sequence i and a three-residue sequence j is shown (the same applies to FIG. 4 described later).

Using the three residue sequence pairs H _{i, j} , A _{i, j} and P _{i, j} , each constraint due to the three-dimensional structure of the protein is expressed as each energy function as shown below.

The energy function U ₁ of the type H is represented by the following equation.
There is at most one residue sequence.

[0034]

(Equation 1) The energy function U ₂ of type A is represented by the following equation. For each three-residue sequence i, a maximum of one three-residue sequence can be paired with type A. This constraint is realized by using the energy function U ₂ of the following equation.

[0035]

(Equation 2) The energy function U ₃ of the type P is represented by the following equation, and the number of the three residue sequences that can be paired with the type P for each three residue sequence i is a maximum of two. This constraint is realized by using the energy function U ₃ of the following equation.

[0036]

(Equation 3) For each three residue sequence i, the maximum number of possible three residue sequences in the β sheet is two, and the energy function U representing the constraint is
₄ is given by

[0037]

(Equation 4) Only one of the maximum, type H, type A and type P is allowed for all three residue sequence pairs,
Energy function U ₅ for that constraint is expressed by the following equation.

[0038]

(Equation 5) In the parallel β sheet, type P is continuous. Therefore, the β-sheet becomes longer when the type P three-residue string pair (i-1, j-1) is adjacent to the type P three-residue string pair (i, j), and the β sheet becomes chemically longer. Become stable. This restriction is realized as follows.

[0039]

(Equation 6) In the antiparallel β-sheet, the binding types of the 3-residue row pairs are alternately arranged as type H-type A-type H-type A-. Accordingly, the type A 3-residue string-3 residue sequence pair (i-1, j + 1) is adjacent to the type H 3-residue sequence-3 residue sequence pair (i, j), and β The sheet becomes longer and chemically stable. The energy function U at this time
₆ is represented by the following equation.

[0040]

(Equation 7) In the anti-parallel β sheet, the binding types of the 3-residue string-3 residue string pairs are alternately arranged like type H-type A-type H-type A-. Prohibit continuation. This constraint is implemented by using an energy function U ₇ follows.

[0041]

(Equation 8) The initial value based on the 3-residue-row-residue-row pair tendency index is almost correct and should not be largely separated. This is (10)
This is realized by Expressions (12) to (12). However, in each formula, H
^{The initial} _{i, j} , A ^initial _{i, j} and P ^initial _{i, j} are the binding energies for all the combinations of the three residue sequences generated by the partial sequence combination generator 13 in FIG. 11 is a calculation result calculated by the binding energy calculation unit 12 using the binding energy stored in 11, and is an initial value given to the Hopfield neural network unit 14.

[0042]

(Equation 9) The Hopfield neural network state energy function E was a weighted sum of these energy functions U _k as shown in the following equation.

[0043]

(Equation 10) Here, in the above equation, α _k is a weight coefficient.

To make the Hopfield neural network operate so as to minimize the energy function E, a steepest descent method is used. Therefore, the energy function E is partially differentiated by H _{i, j,} based on the partial differential coefficients, H _i as shown in the following _equation, the variation in membrane potential of the _{j ΔmH} i, _j
Decide.

[0045]

[Equation 11] Here, the Hopfield neural network state energy function E is expressed as cell activity H _{i, j} , A _{i, j} and P
Since it is almost linear with respect to _{i and j} , the derivation of the operation equation is simple. Using the membrane potential mH _{i, j} of the type H updated based on the change amount ΔmH _{i, j} of the membrane potential _{, the} cell activity H _{i, j} is determined by the following equation.

[0046]

(Equation 12) The membrane potentials mA _{i, j} and mP _{i, j} of type A and type P are also obtained in the same manner as described above, and the activity A of those cells is obtained.
_{i, j} and P _{i, j} are updated based on the partial derivatives.

The above operation formulas and cell activities A _{i, j} and P
By operating the Hopfield neural network unit 14 using the equations (14) and (15) corresponding to _{i and j, the} secondary array structure having a higher reliability than the conventional one in consideration of the three-dimensional structure restriction is considered. Can predict.

Next, a second embodiment of the present invention will be described. The overall configuration of the second embodiment is the same as that of FIG. 1, but the configuration of the Hopfield neural network unit 14 is different. FIGS. 4A and 4B show a part of the Hopfield neural network unit 14 according to the second embodiment. FIG. 4A shows a type H, FIG.
FIG. 3C shows the type P. In the second embodiment, in addition to the configuration shown in FIG. 3, the Hopfield neural network unit 14 includes two bits as shown in FIG.
× (N-1) using a ^2/2 cells.

As shown in FIG. 4, a cell for expressing a cut in a parallel β sheet is denoted by SP, and a cell for expressing a cut in an antiparallel β sheet is denoted by SA. In the second embodiment, using these, the Hopfield neural network unit 14 expresses the start and end of the β-sheet and uses an energy function expressing the constraint on the continuity of the sheet. Therefore, the energy function is additionally changed as follows with respect to the first embodiment.

In the parallel β sheet, a new variable (cell) SP relating to a break is introduced in the energy function for the continuation of the type P. As the nature of this SP,
When SP _{i, j} is “1.0”, for a parallel β sheet,
The three residue sequence pair (i-1, j-1) and the three residue sequence pair (i,
j) is discontinuous, and one of the three residue sequence pairs belongs to the parallel β sheet, and the other does not belong to the parallel β sheet. When SP _{i, j} is “0.0”, it indicates that the sheet is continuous in the sense of a parallel β sheet. Using SP, the continuity of the parallel β sheet is realized by the following equation instead of equation (7).

[0051]

(Equation 13) In the anti-parallel β sheet, the following expression is used instead of expression (8).

[0052]

[Equation 14] The Hopfield neural network state energy function E was a weighted sum of these energies U _k as shown in the following equation.

[0053]

(Equation 15) Here, in the above equation, α _k is a weight coefficient.

In this embodiment, the steepest descent method is used to operate the Hopfield neural network unit 14 so as to minimize the Hopfield neural network state energy function E. For this purpose, the above energy function E is partially differentiated by the activity of each type of cell, and the Hopfield neural network unit 14 is operated based on the partial differential coefficient. It is possible to predict a likely secondary structure sequence. In this embodiment, as shown in FIG. 4, since the cells SA and SP representing the cuts are used, the range of the length of the β sheet can be known more accurately than in the first embodiment. The reliability of the prediction of the substructure sequence can be further improved.

Next, a third embodiment of the present invention will be described. The overall configuration of this embodiment is the same as that of FIG. 1, but the binding energy holding unit 1 holds the binding energy of the α helix in addition to the binding energies of the first and second embodiments. The Hopfield neural network unit 14 expresses the beginning and end of a spiral α-helix and uses an energy function expressing a constraint on the continuity of the α-helix.

The frequency at which the partial sequence of 7 residues appears in the α-helix is determined, the binding energy of the α-helix is calculated based on the occurrence frequency, and the calculated binding energy is stored in the binding energy holding unit 11 of FIG. The number of residues is tentatively determined from the number of the three-dimensional structure database, and partial sequences having other lengths can be used.

In this embodiment, when the partial sequence combination generator 13 shown in FIG. 1 obtains a sequence having a length N and an unknown tertiary structure, N binding bonds to the α helix from the binding energy holding portion 11 are obtained. Energy binding energy calculator 1
2 and calculate the result as the initial value Helix ^initial
_{i, j} are set in the Hopfield neural network unit 14. Here, whether a certain residue is included in the α helix is represented by, for example, a real number of [0.0, 1.0].

In this embodiment, in addition to the configuration of the Hopfield neural network unit 14 of the second embodiment, in order to represent the N cells for the α-helix and β-sheet and the break of the α-helix, N cells are used. FIG. 5 shows a part of the Hopfield neural network unit 14 of this embodiment. In the figure, Sheet _i is a variable indicating the degree of β-sheet, Helix _i is a variable for α-helix, SHeli
x _i is a variable indicating a break in the α helix.

As an energy function used in the Hopfield neural network unit 14 in this embodiment, the following energy function is further added to each energy function of the second embodiment.

Predicted value Helix for α helix
_{i, j} Since most correct order not leave the initial value Helix ^initial _i, and _j increase, using an energy function U ₁₄ follows.

[0061]

(Equation 16) Further, in order to prevent overlap and α-helices and β sheets residue level, using the energy function U ₁₅ follows.

[0062]

[Equation 17] Here, in the above equation, Sheet _i is, as shown in FIG.
This is a variable indicating the degree of the β sheet, and is obtained from the following equation.

[0063]

(Equation 18) In order to make the α helix continuous, an energy function U ₁₆ represented by the following equation is used using SHelix representing a break in the α helix.

[0064]

[Equation 19] Further, the energy function U _{17 of the} following equation is used to make the length of the α helix 4 or more.

[0065]

(Equation 20) Hopfield Neural Network state energy function E, as represented by the following formula, and a weighted sum of the energy function and their respective energy function U ₁₄ ~U ₁₇ used in the second embodiment.

[0066]

(Equation 21) Here, in the above equation, α _k is a weight coefficient.

The Hopfield neural network unit 14 uses the steepest descent method to operate so as to minimize the energy function E. Therefore, the Hopfield neural network unit 14 partially differentiates the Hopfield neural network state energy function E by the activity of the cell, and operates based on the partial derivative. This makes it possible to predict the more likely sequence secondary structure of a protein or RNA sequence whose sequence is unknown, in consideration of the restriction of the three-dimensional structure.

Next, a fourth embodiment of the present invention will be described. FIG. 6 is an overall configuration diagram of a fourth embodiment of an array secondary structure prediction device according to the present invention. In the figure, the same components as those of FIG. 1 are denoted by the same reference numerals, and description thereof will be omitted. As shown in FIG. 6, this embodiment is characterized in that a folding simulation unit 17 for simulating the folding of the three-dimensional structure of the array is provided.

This folding simulation unit 17
Based on the prediction of the α helix and β sheet obtained by the Hopfield neural network unit 14, the molecular model is operated in a computer to simulate whether or not it can be folded. That is, the folding simulation unit 17 receives the value of the operating Hopfield neural network unit 14 as an input, performs folding simulation, and performs feedback on the value of the Hopfield neural network unit 14. The simulation of the folding itself has hitherto been described, for example, in Reference 6 (Hirosawa M., Feldmann R., Rawn D., Ishika
wa M., Hoshida M., and Michaels G., "Folding Simulati
on using Temperature Parallel Simulated Annealin
g ", Proceddings of FGCS 1992, pp. 300-306 (1992)).

According to this embodiment, for example, the residues contained in the β sheet have a sufficient distance between the residues, but the α helix is inserted in the middle, so that It is possible to know in advance that the distance between them becomes shorter, and that they will not be paired in the β sheet.

The present invention is not limited to the above embodiment, but is also applicable to prediction of the sequence secondary structure of other nucleic acid sequences such as DNA (deoxyribonucleic acid) other than RNA.

[0072]

As described above, according to the present invention,
Considering the restrictions of the three-dimensional structure, the amount of calculation for the search increases in proportion to the fourth power of the length as in the conventional case, so that it is difficult to apply to a long array or β There is no possibility that the prediction of the sheet becomes impossible, and the prediction of the secondary structure of the arrangement can be performed more accurately even in the case of a long arrangement or a β sheet at a long distance as compared with the related art.

According to the present invention, folding simulation is performed by receiving the value of the operating Hopfield neural network as an input, and feedback is performed on the value of the Hopfield neural network. Can be improved, and the reliability of predicting the secondary structure of a protein or nucleic acid sequence whose tertiary structure is unknown can be greatly improved.

[Brief description of the drawings]

FIG. 1 is an overall configuration diagram of a first embodiment of a device of the present invention.

FIG. 2 is an explanatory diagram of three types of a 3-residue-string-residue-string pair spanning between strands in a β-sheet of a protein.

FIG. 3 is a diagram illustrating a first example of a Hopfield neural network unit in FIG. 1;

FIG. 4 is a diagram illustrating a second example of the Hopfield neural network unit of FIG. 1;

FIG. 5 is a diagram illustrating a third example of the Hopfield neural network unit in FIG. 1;

FIG. 6 is an overall configuration diagram of a fourth embodiment of the device of the present invention.

[Explanation of symbols]

 DESCRIPTION OF SYMBOLS 11 Binding energy holding part 12 Binding energy calculation part 13 Partial sequence combination generation part 14 Hopfield neural network part 15 Three-dimensional structure unknown sequence 17 Folding simulation part

Continuation of the front page (58) Fields investigated (Int.Cl. ⁷ , DB name) G06N 3/10 G01N 33/48 G01N 33/68 G06F 19/00 110 INSPEC (DIALOG) JICST file (JOIS) WPI (DIALOG)

Claims (6)

(57) [Claims] (1) a method for forming a bond between constituent units of a protein and a nucleic acid sequence;
The input energy is used to determine the unknown protein
All combinations of two subsequences of white matter or nucleic acid sequences
After calculating the binding energy for the
The combination of the sub-arrays with the lowest sum of
Hopfield neural networks expressing body structure constraints
Network, and the input three-dimensional structure is not
Predict sequence secondary structure of known protein or nucleic acid sequenceArray
In the secondary structure prediction method, Obtained by the Hopfield neural network
Whether the result can be folded
Input the combination of foldable sub-arrays
Sequence of protein or nucleic acid sequence of unknown tertiary structure
Predict as structure Sequence secondary structure prediction characterized by the following:
Method. (2)The connection between the constituent units of protein and nucleic acid sequences
A binding energy holding unit that holds the combined energy, and an input
Two parts of a protein or nucleic acid sequence of unknown conformation
Subarray combinations that generate all combinations of minute arrays
And the binding energy from the binding energy holding unit.
From the partial array combination generator using energy
Binding energy for all combinations of two subsequences
A binding energy calculation unit for calculating energy,
The binding energy calculated by the energy calculator is initially
Is set as the value and the lowest
The combination of the partial sequences that generate the binding energy
The input three-dimensional structure of the unknown protein or nucleic acid sequence
Hopfield neural predicting as secondary structure
And a secondary network structure prediction device having
And The binding energy holding unit is a combination of protein partial sequences.
Regarding the β sheet expressed by matching, the parallel type and the
There are three types depending on the parallel type and the hydrogen bonding pattern.
Holding the binding energy of each type
In the energy calculator, the three types of coupling energy
Calculate the binding energy for each of the lugis
To An array secondary structure prediction device, characterized in that: 3. The Hopfield neural network unit is adapted to classify β sheets represented by a combination of partial sequences of proteins into three types according to a parallel type, an antiparallel type, and a hydrogen bond pattern. 3. The apparatus according to claim 2, wherein the apparatus operates using an energy function expressing a constraint. 4. The Hopfield neural network unit represents a start and an end of the β sheet,
The array secondary structure predicting apparatus according to claim 2, wherein the apparatus operates using an energy function expressing a restriction on sheet continuity. 5. The Hopfield neural network unit operates using an energy function that expresses a start and an end of an α-helix expressed by a combination of partial sequences of proteins, and expresses a restriction on continuity of the α-helix. The sequence secondary structure prediction device according to any one of claims 2 to 4 , wherein: 6. A folding simulation unit which receives a value of the Hopfield neural network unit in operation as input, performs a folding simulation, and feeds back the value of the Hopfield neural network unit. The sequence secondary structure prediction device according to any one of claims 2 to 5 .