JP2940529B2 - Protein three-dimensional structure prediction method - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

The present invention relates to a method for predicting the three-dimensional structure of a protein from the amino acid sequence of the protein whose three-dimensional structure is unknown.

[0002]

2. Description of the Related Art As a local three-dimensional structure of a protein,
There are secondary structures such as α-helix and β-sheet, and motifs such as zinc finger and leucine zipper. It is generally considered that if it becomes possible to predict the local structure of a protein with respect to the amino acid sequence of a protein whose structure is unknown, it is possible to predict the three-dimensional structure of the entire protein.

For example, the problem of protein secondary structure prediction is as follows:
This is a problem that has been solved for more than 20 years. Conventionally, each residue in the primary structure of a protein (hereinafter, the residue to be predicted is referred to as a central residue) has 3 (or 4) types. Has been treated as a matter of predicting which of the secondary structures corresponds to As a method for predicting the secondary structure of a protein according to the prior art, for example, a US magazine “Biochemistry” published in 1974 (Biochemistry)
y), Vol. 23, pp. 222-245.
u) and Fasman's dissertation "Prediction of Protein Conformation"
(Prediction of protein co
nformation) (hereinafter abbreviated as CF method), 19
US journal "Journal of Molecule Biology" published in 1978
Cultural Biology) Volume 120 97-12
"Analysis of the Accuracy and Implications of Simple Methods for Predicting the Secondary Structure of Globular Proteins" by Garnier et al.
s of the accuracy and imp
lations of simple method
d for predicting the second
day structure of grobula
r proteins) (hereinafter abbreviated as GOR method), 1
987 US journal "Journal of Mole Biology" (Journal of Mol
volume 198, 425-
Gibrat et al., Pp. 443, "Father Developments of Protein Secondaries Structure Prediction Using Information Theory: New Parameters and Conclusion of Residue Pairs" (Further developments)
of protein secondary tru
cure predicon using in
formation theory: New para
meters and considerationo
fresidue pairs (hereinafter abbreviated as GGR method), a US magazine published in 1988, "Journal of Molecular Biology" (Journal o
f Molecular Biology) No. 202
Vol. 865-884, published by Qian et al., "Predicting the Secondary Structure of Globber Proteins Using Neural Networks Models" (Pred).
icting the second STR
cure of global protein
suspending neural network mod
els) (hereinafter abbreviated as QS), and the 1993 American Society of Hawaii International Conference
Of System Sciences "(Hawaii I
international Conference o
f System Sciences, Proceedings, Vol. 1, pages 659-668, and a paper by Mamizuka et al., "Protein α-Helix Region Prediction Based on Stochastic Rule Learning" (Protein α-Helix Region)
Prediction Based-on Stoc
historical Rule Learning) (hereafter,
MY method).

[0004] In the CF method, the statistical appearance frequency of amino acids in each secondary structure is obtained from a protein structure database, and the frequency table is used to make predictions based on empirical rules. Also, the GOR method calculates the sum of the information amounts independently brought about by the residues several residues away from the secondary structure of the central residue, makes a prediction from the relative value, and calculates the GGR The method predicts the secondary structure of the central residue from the sum of the information content provided by the residue and the residues several residues away from the residue. Furthermore, QS
The method uses a three-layer feed-forward network, inputs a sequence containing eight residues before and after the central residue, and uses a neural network to determine the contribution of the central and peripheral residues to the secondary structure. Prediction is performed by extracting. The MY method calculates the probability that each amino acid is an α-helix at each residue position in the training sequence as a probability distribution, and then calculates the probability of the α-helix for each region of the test sequence.

[0005]

In the amino acid sequence of a protein, it is considered that each amino acid residue retains a dependency and forms a local three-dimensional structure or a functional site. Therefore, in order to represent and predict the local three-dimensional structure of a protein, it is considered important to express the dependence between the residues in the local three-dimensional structure. However, there has been no study on a method for automatically extracting a dependency depending on the residue position in the form of a network or a method for predicting unknown data by using the dependency as a rule. , Such a method was not established.

[0006]

According to a first aspect of the present invention, there is provided a method for predicting a three-dimensional structure of a protein, comprising the step of learning a rule of a protein structure, wherein the step comprises the steps of: As an input, each dependency is a conditional probability, and using a predetermined information criterion, it is said that the dependency does not cycle through other residue positions at which each residue position of the amino acid sequence is dependent. By determining under the restriction, a probabilistic rule expressing the dependency is constructed and output.

According to a second aspect of the present invention, there is provided a method for predicting a three-dimensional structure of a protein, comprising the step of learning the rules of the protein structure, wherein the step comprises inputting a real-valued attribute of an amino acid sequence of training data comprising positive and negative examples. Each dependency is a conditional probability, and using a predetermined information criterion, the other residue positions on which each residue position of the amino acid sequence is dependent are limited by the restriction that the dependency does not circulate. By determining below, a probabilistic rule expressing the dependency is constructed and output.

[0008]

Next, the present invention will be described in detail with reference to the drawings.

FIG. 1 is a flow chart for explaining an embodiment of a protein three-dimensional structure prediction method related to the present invention. In this embodiment, it is assumed that an α-helix is treated as a local three-dimensional structure of a target protein.

In step 11, an amino acid sequence of a protein whose α-helix region is known is aligned with the same family of proteins, for example, the same protein of different species, and the partial sequence corresponding to the α-helix is aligned. Is extracted as a positive example of an α helix.

For example, in the case of a protein β-chain called hemoglobin, the position of the α-helix of human hemoglobin has been clarified from the result of X-ray crystal diffraction.
It is known to have a region of eight α-helices. Therefore, human hemoglobin β chain is aligned with other species, for example, chimpanzee, horse and other species of hemoglobin β chain, and a region corresponding to eight α helices is defined as a positive example of α helix Extract.

In step 12, each sequence of the amino acid sequence database with known α-helix position is aligned with the partial sequence corresponding to α-helix of the protein with known α-helix position, Non-corresponding subsequences are extracted as a negative example relative to the positive example of the α helix extracted in step 10.

In the example of the hemoglobin β chain, the partial sequence corresponding to the eight α helices is, for example, PDB
Alignment is performed for several sequences in a protein structure database such as (Data Protein Bank), and in each partial sequence obtained as a result of the alignment, if the structure of the sequence is not α-helix, these are regarded as negative examples. Extract as For example, in the alignment at the time of extracting a negative example, it is conceivable that a partial sequence having homology of a certain ratio or more is regarded as a negative example. Specifically, there is a method using a partial sequence having a homology of 30% or more by alignment as a negative example.

For the number of data to be extracted, for example, α
It is conceivable that the ratio of the positive example to the negative example in each region serving as the positive example of the helix is made equal in each region. Furthermore, more specifically, it is conceivable that the same number of positive examples and negative examples is used as the ratio.

Step 13 is a step of calculating real-valued parameters of the probabilistic rule from the positive and negative examples extracted in steps 11 and 12.

The stochastic rule here is a probability distribution that gives a probability that an α-helix corresponds to an arbitrary region of a given sequence. Each X _i (i =
1,. . . , N) are attribute value spaces, and X is their direct product, that is, X = X ₁ × X ₂ ×. . . × X
Write _n .

For example, X represents one set of 20 amino acids, or X = X ₁ × X ₂ and X
₁ there is a case where numeric range range and X ₂ numbers expressing hydrophobicity indicates the molecular weight.

For a window W of length L in the positive example of the α helix, a region S of arbitrary length L in the test sequence is W
The likelihood corresponding to the part is determined as follows. First, X _t (hereinafter referred to as a variable) is the t-th residue position counted from the left of the sequence S, and π _t is defined as X _{t in the} region W.
There set of dependent residue position (hereinafter, the set of parent variables of [pi _t X _t, the X _t is referred to as [pi _t of the child variable) to. here,

[0019]

(Equation 1)

Let the conditional probability that X _t depends on π _t be in the region W, and assume that the probability Pw (S) that the region S corresponds to the W portion can be written as follows.

[0021]

(Equation 2)

The right side of Equation 1 corresponds to a network structure by using variables as nodes and extending an arc from a parent to a child between the variables. For example, if the region S is composed of three residues, and the connection probability of each residue in the region S can be specifically written as the following expression, the following expression corresponds to the network of FIG.

[0023]

(Equation 3)

Further, each

[0025]

(Equation 4)

Is determined from the given case data including the positive and negative examples, for example, as follows.

[0027] First, in the t-th residue position, partial area not overlapping the possible range of real-valued attributes (hereinafter, referred to as cells) finite division, the total number of cell m _t to, C _i Is the i-th cell.

[0028] If the residue of the t-th position is included in C _i of the m _t number of cells, occurrence of X _t probability P (X _t
= I) = p _i (t). here,

[0029]

(Equation 5)

This is called a probability parameter. FIG. 4 shows an example of the structure of the finite division, in which the value is from 0 to 1.
Shows a case where a probability parameter is estimated by one attribute having a range of.

The probability parameter is estimated using the number of positive and negative examples contained in each cell.

[0032]

(Equation 6)

Is the number of positive examples included in the i-th cell at the t-th position, and N _i (t) is the sum of the number of positive examples and the number of negative examples included in the i-th cell at the t-th position. , Let the estimated value at the i-th cell at the t-th position be p _i (t).
For example, a probability parameter for each cell is calculated by the Laplace estimator of the following equation.

[0034]

(Equation 7)

However, not only the Laplace estimator but also many estimators can be used. Next, similarly, X _t
And the joint probability P (X _{t, π t)} of [pi _t can be calculated by using the estimated amount. For example, if the element of π _t is only the variable X _s , then X _s is the same as X _t and m _s
And finite divided into individual partial areas, s-th residue is included in C _j in m _s-number of cells,

[0036]

(Equation 8)

And the probability parameters p _i , j (t,
s) is estimated.

At each of the t-th and s-th positions, based on the number of positive and negative data contained in each cell,

[0039]

(Equation 9)

Is the number of positive examples contained in the i, j-th cell at each position, and N _i , j (t, s) is the number of positive examples and the negative example contained in the i, j-th cell at each position. The sum of numbers. From this, for example, a probability parameter is estimated by the following Laplace estimator.

[0041]

(Equation 10)

[0042] Finally, using these estimated probability parameters, Xt conditional probability of under that π _t is present

[0043]

[Equation 11]

Is calculated as a probability parameter. As described above, [pi elements _t is only variable X _s, t th and s-th each m _t pieces each position, and finite divided into m _s pieces, further residue cell C _i of each position , C _j ,

[0045]

(Equation 12)

And the probability parameter

[0047]

(Equation 13)

Is calculated as follows:

[0049]

[Equation 14]

In step 14, the structure of the stochastic rule is determined. That is, for each variable X _t , a set of parent variables π _t is determined using the information criterion.

Hereinafter, the minimum description length (Mi
minimum Description Length
(MDL) An example of a network configuration method when applied to the principle (hereinafter, the MDL principle) will be described. In addition,
For the MDL principle, see the 14th edition of the American magazine "Automatica" published in 1978.
Volume 465-471, Rissane.
n) "Modeling by Showtest Data Description" (Modeling by
shortest data description
n).

According to the MDL principle, the rule that minimizes the sum of the data description length calculated from given case data and the rule description length is determined as the optimal rule. Therefore, the data description length of the probabilistic rule and the rule description length are calculated from the positive example, the negative example, and the real value parameters obtained in steps 11, 12, and 13.

In the example described here, it is considered that, for each residue position, a position that is dependent on the position is determined. That is, the parent variable is determined independently for each variable.

First, attention is paid to the t-th residue position. The dependency between the variable X _t and its parent variable set π _t is determined by the conditional probability

[0055]

(Equation 15)

Is represented by Here, the number of parent variables is k,
The residue position of the parent variables and t _k from t _l in order, also,
t and t _l from t _k each m _t the total number of cells at residue position of, m _tl, ···, and m _tk. Furthermore, residues of the variable X _t is i, the residues X _tk from the variable X _tl,
Each j _l, ···, a positive number of cases that have attributes such as contained in j _k th cell

[0057]

(Equation 16)

[0058], positive sample that residues of the variable X _t is i, residues X _tk from the variable X _tl is respectively have j _l, · · ·, attributes such as contained in j _k th cell N _{i, jl, · ・ ・ , Jk} (t, t _l , ...
·, T _k), conditions residues of the variable X _t is i, residues X _tk from the variable X _tl is respectively have j _l, · · ·, attributes as fall j _k th cell With the probability parameter

[0059]

[Equation 17]

It is assumed that:

Then, the data description length of the probabilistic rule here is given by the following equation by taking the log likelihood of the rule negative from the probability parameter calculated in step 13.

[0062]

(Equation 18)

Further, the description length of the rule of the probabilistic rule is given by the following equation.

[0064]

[Equation 19]

[0065] Therefore, with respect to the variable X _t, which corresponds to t-th residue positions, by selecting the set of parent variables minimizing the sum of (6) and (7), stochastic The structure of the rule is determined.

In step 15, using the probabilistic rule constructed in step 14, the activity of each region of the given test data array is calculated.

Here, likelihood is used as the activity.

First, an area W having an arbitrary length L in the positive example of the training sequence is extracted. The parent variable of the variable corresponding to each residue position of W has been determined in step 14. Further, the real-valued parameter of the conditional probability representing the dependency between each variable and its parent variable is calculated in step 13.

Next, this area W is applied to a partial sequence S of an arbitrary length L of the test sequence, and the likelihood of the α helix is calculated.

For example, in the area W where L = 3, (2)
The structure of the stochastic rule such as the equation is determined, and in the region S of L = 3 of the test sequence, each residue in the region is 2,
Suppose that it has real-valued attributes that fall into the first and third cells. Then, the likelihood that this area S is W can be calculated by the following equation.

[0071]

(Equation 20)

This operation is performed on all the sub-arrays of length L in the test array using a probabilistic rule composed of all possible length L regions in the positive example of the training array.

In step 16, for a region S having an arbitrary length L in the test sequence, a plurality of likelihoods calculated by all possible length L regions of the positive example in the training sequence are used. , The maximum likelihood is selected as the α-helix likelihood of the region S.

The operations in steps 15 and 16 can be summarized as follows. That is, all sets of the length L region in the positive sequence in the training sequence are A, and the likelihood P (S) of the α helix with respect to the partial sequence S having the length L of the test sequence is calculated by the following equation.

[0075]

(Equation 21)

By repeating this operation for each area of the test array, the likelihood of the α helix is calculated for each area of the test array.

Here, if there are a plurality of α-helix regions, the same likelihood calculation is performed for each α-helix region, and the maximum likelihood is selected as the optimum value over the entire α-helix region.

Further, in each of the regions to which the likelihood in the test sequence is given, the maximum likelihood is set to the optimum value of each residue in the region. The average of the obtained likelihood of the region containing the residue is used as the optimum value of each residue, and the like, and the likelihood change for the entire test amino acid sequence is output.

The learning and predicting method in FIG.
The present invention can be applied to feature extraction and prediction of local regions such as secondary structures and motifs other than the α-helix. FIG. 2 is a flowchart illustrating an embodiment of the protein three-dimensional structure prediction method of the present invention. In the present embodiment, it is assumed that an α-helix is used as a target secondary structure.

In step 21, the same processing as in step 11 in FIG. 1 is performed to extract a positive example required for α-helix region prediction.

In step 22, the same processing as in step 12 in FIG. 1 is performed to extract a negative example necessary for α-helix region prediction.

In step 23, the same processing as in step 13 in FIG. 1 is performed, and the real-valued parameters of the probabilistic rule are estimated from the positive and negative examples extracted in steps 21 and 22.

Step 24 is a step of constructing a probabilistic rule as a constrained probabilistic rule as a joint probability distribution of all variables in the local structure region, with restrictions imposed on the construction of the probabilistic rule. Included in the second aspect. The restriction here is a restriction in which the direction of the arc is non-circular so that the probability distribution does not conflict when the probabilistic rule is represented by a network structure as shown in FIG. For example, FIG. 3 is an example of a non-circulating network, in which X ₁
If the arc extending from to X ₃ is extended from X ₃ to X ₁ , the network becomes a circular network and the creation of such a network is not allowed.

As a method of adding a restriction, for example, a method in which each variable is ordered and only variables having a small order can be used as a parent variable, or a case where a dependency relation in which an arc circulates is formed. Only, a method of preventing the dependency from being established can be considered.

In step 25, the same processing as in step 15 in FIG. 1 is performed, and the probabilistic rule whose structure has been optimized using step 24 is used. Calculate the degree.

In step 26, the same processing as in step 16 of FIG. 1 is performed, and a change in the activity for the entire array is output from the plurality of activities determined in step 25.

The learning and predicting method in FIG.
The present invention can be applied to feature extraction and prediction of local regions such as secondary structures and motifs other than the α-helix.

[0088]

According to the present invention, from the amino acid sequence information of a protein having a known three-dimensional structure, the local three-dimensional structure of a protein whose local three-dimensional structure is unknown can be predicted with higher accuracy than the prior art. For example, the MY method, which is one of the conventional methods, does not consider the dependency between the residue positions in the local region at all. However, due to the configuration of the probabilistic rule reflecting the dependency between the residue positions, The feature extraction and prediction of the local region with higher accuracy can be performed. In addition, the optimization based on the information criterion makes it possible to theoretically optimize the structure of the stochastic rule.

[Brief description of the drawings]

FIG. 1 is a flowchart showing one embodiment of a method for predicting a protein three-dimensional structure related to the present invention.

FIG. 2 is a flowchart showing an embodiment of the protein three-dimensional structure prediction method of the present invention.

FIG. 3 is a schematic diagram showing a dependency relationship between variables of a stochastic rule used in the present invention.

FIG. 4 is a schematic diagram showing a specific example of finite division performed at each residue position in the present invention.

[Explanation of symbols]

 11 Extraction of positive examples 12 Extraction of negative examples 13 Estimation of real-valued parameters 14 Determination of the structure of probabilistic rules 15 Calculation of activity for test sequences 16 Calculation of predicted values for test sequences 21 Extraction of positive examples 22 Extraction of negative examples 23 Real-valued parameter estimation 24 Probability Determining the structure of a statistical rule 25 Calculating activity for a test sequence 26 Calculating a predicted value for a test sequence

──────────────────────────────────────────────────続 き Continued on the front page (58) Field surveyed (Int.Cl. ⁶ , DB name) G06F 17/30 G06F 17/00 JICST file (JOIS)

Claims (2)

(57) [Claims] 1. A method for predicting a three-dimensional structure of a protein, comprising the step of learning a rule of a protein structure, wherein the step receives an amino acid sequence of training data comprising a positive example and a negative example and determines each dependency with a conditional probability. By using a predetermined information amount criterion, by determining the other residue positions on which each residue position of the amino acid sequence is dependent under the restriction that the dependency does not circulate, A protein three-dimensional structure prediction method, comprising constructing and outputting a probabilistic rule expressing a dependency. 2. A method for predicting a three-dimensional structure of a protein, comprising the step of learning a rule of a protein structure, wherein the step comprises inputting a real-valued attribute of an amino acid sequence of training data comprising a positive example and a negative example; Is defined as a conditional probability, and using a predetermined information criterion, the other residue positions on which each residue position of the amino acid sequence depends are determined under the restriction that the dependency does not circulate. A method for predicting a protein three-dimensional structure, wherein a probabilistic rule expressing the dependency is constructed and output.