JP2551297B2 - Protein three-dimensional structure prediction method - Google Patents

Description

Detailed Description of the Invention

[0001]

TECHNICAL FIELD The present invention relates to a method for predicting the tertiary structure of a protein from the amino acid sequence of the protein of unknown structure.

[0002]

2. Description of the Related Art One of the problems in predicting the three-dimensional structure in a protein using the amino acid sequence information of the protein is the protein secondary structure prediction problem. The secondary structure refers to a cohesive structure inside the protein such as α helix and β sheet, and the secondary structure prediction problem is
Using the amino acid sequence information of the protein, one of the 3 (or 4) types of secondary structures corresponding to each residue of the primary sequence (hereinafter, the residue to be predicted is the central residue) It is a problem of predicting the secondary structure, and it is considered that the prediction of the secondary structure also enables the prediction of the three-dimensional structure of the protein. FIG. 3 is a schematic diagram showing a method for predicting a secondary structure (α-helix) region of the present invention. As a method for predicting a secondary structure of a protein according to a conventional technique, for example, a US journal “Bio Second of "Chemistry" (Biochemistry)
Vol. 3, pages 222-245, Chou and Fasman, "The Prediction of Protein Conformation" (Predi).
action of protein conforma
(hereinafter abbreviated as CF method), an American magazine “Journal of Molecular Biology” published in 1978 (Journal of Molecular B).
(Analysis of the Accuratey and Implications of Simple Method for Predicting The Secondary Structure of Globler Proteins), published by Garnier et al.
cure and implications o
f simple method for predi
cinging the secondary struc
pure of global protein
s) (hereinafter, abbreviated as GOR method), an American magazine “Journal of Molecular Biology” issued in 1987 (Journal of Molecular Biology).
Biography, Vol. 198, pp. 425-443, Gibrat et al., "Father Developments of Protein Secondaries Structure Prediction Educating Information Theory: New Parameters and Considation of Residue Pairs" (Further develo).
claims of protein second
ry structure prediction u
sing information theory: N
ew parameters and conside
relation of pairs (hereinafter abbreviated as GGR method) and the American magazine “Journal of Molecular Biology” published in 1988.
(Journal of Molecular Bio
202, pp. 865-884, by Qian et al., "Predicting the
Secondary Structure of Globler
Proteins Using Neural Network Models "(Predicating the seco
ndary structure of globul
ar proteins using neural
network models) (hereinafter abbreviated as QS method). The CF method obtains statistical occurrence frequencies of amino acids in each secondary structure from a protein structure database, and uses this frequency table 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, the QS method uses a feedforward type network of three layers, inputs a sequence including 8 residues before and after the central residue, and uses the neural network to make contributions from the central and peripheral residues to the secondary structure. The prediction is performed by extracting with.

[0003]

Prediction for selecting a secondary structure corresponding to each residue of an amino acid sequence from among three types of secondary structures is called three-state prediction, which is a measure of the prediction result. The rate is in the range of 60% in three-state prediction in any of the conventional techniques, and a prediction method with a higher prediction rate is desired only for α-helix. Further, the conventional prediction result is a residue-corresponding prediction that predicts the secondary structure corresponding to each central residue in the amino acid primary sequence, such as which region in the primary sequence corresponds to which secondary structure. Although it is important to make predictions corresponding to regions, such prediction methods have not been sufficiently studied. Furthermore, not only the amino acid sequence as a character string, but also the prediction method by performing the prediction considering the properties of the amino acid (hydrophobicity, molecular weight, etc.) has not been established at all.

[0004]

The first invention is to predict not only a protein of known structure but also a structure of unknown structure in order to predict the structure of the protein from the amino acid sequence of the protein .
The sequence corresponding to a certain three-dimensional structure using the protein
An array that is known to be positive and not positive
A training data extraction system that extracts training data consisting of
Step a, the learning step of performing learning of probabilistic rules from training data consisting of positive cases and negative cases, the learned stochastic rules for each region of the test amino acid sequence data using
And a prediction step for predicting the structure .

A second invention is the training data extraction step.
Is the structurally known protein amino acid sequence,
Aligning proteins that belong to the same family, extracting the partial sequence corresponding to the secondary structure region that is the target of prediction as a positive example of the secondary structure region, and using it as the target of prediction of proteins of known structure Align each sequence in the database consisting of proteins of known structure with the partial sequence corresponding to the secondary structure, and extract the partial sequence that does not correspond to the secondary structure to be predicted as a negative example of the secondary structure region. And a step.

In a third aspect of the invention, the learning step estimates a real-valued parameter of the stochastic rule by using a stochastic rule from the types of amino acids in the learning data consisting of the positive example and the negative example. It is characterized by doing.

In a fourth aspect of the present invention, the learning step uses a probabilistic rule from the real-valued attributes of amino acids in the learning data consisting of the positive example and the negative example, whereby the real-valued parameter of the probabilistic rule is obtained. Is estimated.

In a fifth aspect of the present invention, the learning step uses a probabilistic rule from the real-valued attributes of amino acids of the learning data consisting of the positive example and the negative example, whereby the real-valued parameter of the probabilistic rule is obtained. estimating a, the
Optimizing the model in the probabilistic rule using an information criterion.

[0009] A sixth invention is the prediction step, the
Using the probabilistic rules learned by the learning step, before
It is characterized in that it comprises a step of calculating the activity level of each region of the test amino acid sequence data and a step of selecting an optimum value from the calculated activity levels.

[0010]

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

FIG. 1 is a flow chart for explaining an embodiment of the protein tertiary 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.

Step 10 is included in the second invention.
In this step, the proteins of the same family, for example, the same proteins of different species are aligned with the amino acid sequence of the protein whose α-helix region is known, and the partial sequence corresponding to the α-helix is Extract as a positive example of a helix.

For example, in the case of the β chain of a protein called hemoglobin, the position of the α helix of human hemoglobin has been clarified from the results of X-ray crystal diffraction,
It is known to have a region of 8 α-helices. Therefore, the human hemoglobin β chain is aligned with the hemoglobin β chain of other species such as chimpanzee and horse, and the region corresponding to 8 α helices is used as a positive example of α helix. Extract.

Step 20 is included in the second invention.
In this step, for the partial sequence corresponding to the α-helix of the protein whose α-helix position is known,
Aligning each sequence in the amino acid sequence database of known α-helix positions, a partial sequence not corresponding to the α-helix is extracted as a negative example for the positive example of the α-helix extracted in step 10.

In the example of the hemoglobin β chain, for the partial sequence corresponding to the 8 α helices, 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 to make the ratio of positive and negative examples in each region that is a positive example of the helix the same in each region, and for example, it is possible to make the same number of positive and negative examples.

Step 30 is included in the third invention, the fourth invention, and the fifth invention in common, and is a step of estimating the real-valued parameter of the stochastic rule. In this step, the real-valued parameter of this probabilistic rule is estimated by using the probabilistic rule from the learning data consisting of the positive example obtained in step 10 and the negative example obtained in step 20. The structure of the probabilistic rule in this step is shown below.

The probabilistic rule is a probability distribution which gives a probability that an α helix corresponds to an area of an arbitrary given array. Each χ _i (i = 1, ...,
Let n be the space of attribute values, and let χ be the direct product of them, that is, χ = χ ₁ × χ ₂ × ... × χ _n .

For example, χ represents a set of 20 kinds of amino acids, or χ = χ ₁ × χ ₂ , where χ ₁
There is a case where is a range of numerical values showing hydrophobicity and χ ₂ is a range of numerical values showing molecular weight. The former case in this example is used in the third invention, and the other cases are used in the fourth invention and the fifth invention. S is an array having a length W of a certain region, each S is regarded as an element of χ × χ × ... × χ, and X _i is the i-th residue counted from the left of the sequence S,
Let P (α | X _i ) be the probability that the secondary structure corresponding to X _i is an α helix. Here, the probability P (α | S) that the secondary structure corresponding to the sequence S is an α helix is P (α
It is assumed that the product of | X _i ) is multiplied by

P (α | S) = П _{i = 1} ^w P (α | X _i ) Further, as a concrete expression of each P (α | X _i ), for example, a finite division stochastic rule is used. The finite division stochastic rule is a conditional probability distribution having the following structure, and is constructed as follows. The possible range of the real value of the attribute at the i-th residue of the array S is divided into non-overlapping partial regions (hereinafter referred to as cells), m is the total number of cells, and C _k is the k-th cell. Then, when X _i is included in C _k of m cells, P (α | X _i ) = P
_{Let k} (i). here,

[0021]

[Equation 1]

Which is called a probability parameter. FIG. 4 is a schematic diagram showing the structure of the finite division type probability rule. In this figure, as an example, a case where the probability parameter is estimated by one attribute whose value ranges from 0 to 1 is shown.

The probability parameter is estimated using the number of positive and negative examples of data included in each cell. m is the number of cells,
Positive sample number contained N _k ⁺ (i) to k th cell in the i-th position, N _k ^- negative sample number contained (i) to k th cell in the i-th position, N _k (I) k at the i-th position
The sum of the number of positive examples and the number of negative examples contained in the th cell, and the estimated value in the k th cell at the i th position

[0024]

[Equation 2]

It is assumed that For example, a probability parameter for each cell is calculated by the Laplace estimator of the following equation.

[0026]

(Equation 3)

However, not only the Laplace estimator but also many estimators can be used.

Step 40 is included in the sixth invention.
This step uses the probabilistic rule learned in step 30 to calculate the activity for each region of the test data array.

Here, likelihood is used as the activity.

Specifically, consider an α-helix region having a length w for which a stochastic rule is constructed. For the amino acid sequence of the test data, all w-t + 1 partial regions having a length t smaller than the length w of the region are provided, and this w-
Each of the t + 1 partial regions is applied in order from the left of the test amino acid sequence, and the likelihood of each region of the test sequence is calculated.

Now, for the k-th sub-region of length t, the probability parameters of the α-helix region are arranged in order from the left, ξ _k = (θ ₁ , ..., θ _t ), θ _i = (P
₁ (i), ..., P _m (i)) (i = 1, ..., t).

Here, m is the number of cells, and θ _i has already been obtained by learning.

Corresponding to the positions of the w-t + 1 partial areas, the mt-dimensional parameters are obtained in the quantity of _{w-t + 1,} and they are defined as ξ ₁ , ..., ξ _{w -t +1} .

Using the above parameters, an arbitrary length t
With respect to the test amino acid sequence Γ of, the likelihood can be calculated as follows with w−t + 1.

P (α | Γ: ξ _k ) (k = 1, ..., w−t + 1) However, for each k, P (α | Γ: ξ _k ) = Π _{i = 1} ^t P (α | Γ: θ _i ) Here, P (α | Γ: θ _i ) is calculated as P _l (i) (l = 1, ..., M) if X _i enters the l-th cell. Also, P _I (i) (l = 1, ..., M) has already been learned.

For example, in the case where the space of the attribute values of amino acids in the finite division stochastic rule consists of only one attribute value, the number of cells is 3, and the estimator entering the cell is the Laplace estimator. Let's ask by. At this time, some α
A positive number of cases of l th cell in the i-th position of the k-th long 5 subregion of helical regions N _l ⁺ (i), a negative number of cases of l th cell N _l ^- (i) , N ₁ (i) is the sum of the number of positive examples and the number of negative examples. Then, the estimator of the l-th cell at the i-th position is, for example, P _l (i) = (N _l ⁺ (i)
+1) / (N _l (i) +2).

Here, suppose that the test is performed on a region Γ of the window size 5 of the test amino acid sequence, and the region Γ
It is assumed that each residue has a real value of the attribute that enters each of the 1, 2, 3, 2, and 1st cells in the learning rule configured in the partial region. Then, the activity of the region Γ of the test amino acid sequence by the k-th partial region is the likelihood P (α
| Γ: k) is calculated as follows. P (α | Γ: k) = {(N ₁ ⁺ (1) +1) / (N
₁ (1) +2)} {(N ₂ ⁺ (2) +1) / (N
₂ (2) +2)} {(N ₃ ⁺ (3) +1) / (N
₃ (3) +2)} {(N ₂ ⁺ (4) +1) / (N
₂ (4) +2)} {(N ₁ ⁺ (5) +1) / (N
₁ (5) +2)} This likelihood calculation is performed for all the regions that the test amino acid sequence can take, using w−t + 1 partial regions.
If there are a plurality of α-helix regions, the same likelihood calculation is performed for each region.

Therefore, the likelihood is obtained as an output for all regions corresponding to the window size in the test amino acid sequence.

By the likelihood calculation corresponding to each region using the above windows, the probability that the α helix corresponds to each residue is not calculated, but α is calculated for each partial region of the test amino acid sequence. The probability that the helix corresponds can be calculated as the likelihood.

Step 50 is included in the sixth invention.
In this step, one of the plurality of activities calculated in step 40 is found to be the optimum one for Γ, and the change in the activity of the entire test amino acid sequence is output.

Continuing from step 40, the likelihood is used here as the activity.

For example, the optimum value P (α | Γ: ξ _k ^* ) is determined as follows. P (α | Γ: ξ _k ^* ) = max {P (α | Γ: ξ ₁ ),
…, P (α | Γ: ξ _{w -t + 1} )}. If there are a plurality of α-helix regions, a similar likelihood calculation may be performed for each region for the same Γ, and the maximum likelihood may be selected as the optimum value throughout the α-helix region.

Further, in each region given a likelihood in the test sequence, the maximum likelihood is set to the optimum value of each residue in the region, or for each residue in the region, A method such as using the average of the obtained likelihoods of the region containing residues as the optimum value of each residue is used to output the change in the likelihood for the entire test amino acid sequence.

The learning and prediction method shown in FIG.
It can be applied to secondary structure prediction other than α-helix.

FIG. 2 is a flow chart for explaining an embodiment of the protein tertiary 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 60, the same process as in step 10 of FIG. 1 is performed to extract a positive example required for α-helix region prediction.

In step 70, the same processing as in step 20 of FIG. 1 is performed to extract a negative example required for α-helix region prediction.

In step 80, the same process as in step 30 of FIG. 1 is performed to apply the stochastic rule, and the real-valued parameters of this stochastic rule are estimated.

Step 90 is included only in the fifth invention. In this step, the stochastic rule model is optimized using the information criterion. The information criterion to use is
For example, MDL (minimum descriptor)
on length) criteria and the like.

Applying to the finite division type probabilistic rule, the MDL principle is that the optimum probabilistic rule is constructed when the sum of the data description length and the description length by the finite division type probabilistic rule is the minimum. To do. Regarding the MDL principle, the American magazine "Automatic" published in 1978.
(Automatica) Vol. 14, pp. 465-471, by Rissanen, "Modeling by short test data description" (Modeling by shorttest).
data description).

The description length of the data for the finite division type probabilistic rule is obtained by taking the negative of the log-likelihood of the rule.
It can be calculated as follows. However, in the following, the base of the logarithm is all 2.

[0052]

[Equation 4]

The description length of the finite division type stochastic rule is such that the estimated value of each probability parameter p _K (i) is approximately lo.
Since it can be described by g N _k (i) bits, it can be calculated as follows.

[0054]

(Equation 5)

Therefore, according to the MDL principle, the size of the number of cells m that minimizes the following equation is set as the optimum number of cells forming the probability rule.

[0056]

(Equation 6)

The step 100 performs the same process as the step 40 of FIG. 1, and uses the stochastic rule for which the model is optimized using the step 90, for each region of the test amino acid sequence data. Calculate the degree.

In step 110, the same process as in step 50 of FIG. 1 is performed, and a change in the activity for the entire array is output from the plurality of activities obtained in step 40.

The learning and prediction method shown in FIG.
It can be applied to secondary structure prediction other than α-helix.

[0060]

The secondary structure of a protein of unknown secondary structure can be predicted with high accuracy from the amino acid sequence information of the protein of which secondary structure is known. In particular, the α-helix region of albumin can be predicted with high accuracy of 70% or more. Further, by optimizing the model according to the information amount criterion such as the MDL principle, it becomes possible to theoretically optimize the structure of the stochastic rule.

[Brief description of drawings]

FIG. 1 is a flow chart showing an example of a method for predicting a protein three-dimensional structure of 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 of a secondary structure (α helix) region prediction method of the present invention.

FIG. 4 is a schematic diagram showing a specific example of a finite division type probability rule which is an example of a probability rule used in the present invention.

[Explanation of symbols]

 10 Positive Example Extraction 20 Negative Example Extraction 30 Real Value Parameter Estimation by Probabilistic Rule 40 Activity Calculation for Each Region of Test Sequence 50 Prediction Value Calculation for Test Sequence 60 Positive Example Extraction 70 Negative Example Extraction 80 Real Value Parameter Estimation by Probabilistic Rule 90 Optimization by information criterion 100 Activity calculation for each region of test sequence 110 Calculation of predicted value for test sequence

Claims (6)

(57) [Claims] 1. In order to predict the structure of a protein from the amino acid sequence of the protein, all proteins of known structure are used.
Without using a protein of unknown structure,
You can see that it is a positive example that is an array corresponding to the structure and that it is not
Extract training data consisting of
Training data extraction step, a learning step of learning probabilistic rules from the training data consisting of the positive example and the negative example, and a structure of each region of the test amino acid sequence data is predicted using the learned probabilistic rules. A method for predicting a three-dimensional structure of a protein, comprising: 2. The training data extraction step aligns proteins belonging to the same family with an amino acid sequence of a protein of which structure is known , and extracts a partial sequence corresponding to a secondary structure region to be predicted. , Extracting as a positive example of the secondary structure region, and aligning each sequence of the database consisting of the protein of known structure with the partial sequence corresponding to the secondary structure to be predicted of the protein of known structure, The method for predicting protein tertiary structure according to claim 1, comprising a step of extracting a partial sequence that does not correspond to the secondary structure to be predicted as a negative example of the secondary structure region. 3. The learning step estimates a real-valued parameter of the probabilistic rule by using a probabilistic rule from the types of amino acids in the learning data consisting of the positive example and the negative example. The protein tertiary structure prediction method according to claim 1. 4. The estimating step estimates a real-valued parameter of the probabilistic rule by using a probabilistic rule from the real-valued attributes of amino acids in the learning data consisting of the positive example and the negative example. The protein three-dimensional structure prediction method according to claim 1, wherein 5. The step of estimating a real-valued parameter of the stochastic rule by using a probabilistic rule from the real-valued attributes of amino acids of the learning data consisting of the positive example and the negative example in the learning step. And a step of optimizing the model in the stochastic rule by using an information criterion.
The method for predicting protein three-dimensional structure according to 1. 6. The predicting step uses the probabilistic rule learned by the learning step , and for each region of the test amino acid sequence data ,
The method for predicting protein three-dimensional structure according to claim 1, comprising a step of calculating the activity level and a step of selecting an optimum value from the calculated activity levels.