CN105740646A - BP neural network based protein secondary structure prediction method - Google Patents

Abstract

The invention belongs to the field of protein secondary structure prediction methods, relates to a BP neural network training and prediction method used for protein secondary structure prediction and solves the problem of bad prediction effect of the protein secondary structure. The BP neural network training and prediction method comprises the steps of firstly selecting a group of training sample sets with [alpha]-helix, [beta]-sheet and coiling structures accounting for normal proportions from PDB, coding an amino acid sequence of a protein and regarding the coded amino acid sequence of the protein as a network input, and regarding a secondary structure of the corresponding amino acid as a network output; optimizing based on a gradient method, introducing a learning rule which is attached with a momentum item and a self-adaptive learning rate to avoid an oscillation phenomenon and prevent from being trapped in a local minimum value; adopting a six-bit input coding way and a sliding window technology in an input layer, setting a hidden layer structure based on an experience formula and the size of a sliding window; and outputting and predicting classification of the protein secondary structure by an output layer based on a DSSP algorithm.

Description

A kind of secondary protein structure prediction method based on BP neutral net

Technical field

The present invention relates to bio information field, particularly to secondary protein structure prediction method.

Background technology

The higher structure of protein determines its biological function, and wherein secondary protein structure is determined by its aminoacid sequence feature, is the basis of prediction higher structure.Secondary protein structure mainly utilizes amino acid sequence information, analyzes by experiment or Secondary structure is predicted by statistical method, when the accuracy of prediction reaches more than 80%, it is possible to describe the space structure of protein more accurately.Development along with DNA analysis and sequencing technologies, by the derivation of DNA and analysis are obtained substantial amounts of protein sequence information, but the protein structure obtained by traditional experiment and statistical method analysis is little, protein structural database updates slowly, hindering protein structure and the development of function correlational study, the emphasis studied at present starts to turn to gradually finds new Structure Prediction Methods.

Artificial neural network (ArtificialNeuralNetwork, i.e. ANN), abstract by human brain neuroid, from the angle of information analysis and process, the characteristic of natural neutral net is carried out abstract and simulation, by different rules and connected mode network consisting.BP neural network algorithm is one of current most widely used neural network model, is successfully applied in the research in the fields such as information, biology and medical science.The introducing of BP neural network algorithm, makes protein structure prediction obtain new approach.

Summary of the invention

The present invention is directed to the defect that protein secondary structure prediction accuracy rate is low and BP neutral net exists, the problem that the learning process of network is improved, it is provided that a kind of BP neural metwork training for protein secondary structure prediction and Forecasting Methodology.

The present invention is achieved by following proposal, a kind of secondary protein structure prediction method based on BP neutral net, adopts following step to realize:

Step one, choose from PDB one group of α spiral, β-pleated sheet and curling three class formations account for normal rates protein structural database composition training sample set；

Step 2, adopt the amino acid sequence information of six input coding protein, adopt the selected input range of sliding window；

Step 3, BP Neural Network Training Parameter initialize, sliding window position initialization；

The output of step 4 one, hidden layer calculates, and chooses training sample input network, calculates hidden layer output；

The output of step 4 two, output layer calculates, and utilizes hidden layer to calculate the output obtained, and calculates output layer output；

Step 4 three, output layer weights Error Calculation, the actual output of comparing cell and desired output, calculate output layer weights error；

Step 4 four, hidden layer weights Error Calculation, utilize output layer correction error and hidden layer output, calculate and obtain hidden layer weights error；

Neuronic weights and threshold value, according to Error Calculation weighed value adjusting amount, are adjusted by each neuronic weights in step 4 five, reversely adjustment network；

Step 4 six, sliding window move to right one, skip to 41, until whole protein sequence has calculated；

Step 4 seven, select next training sample to learn, skip to 41, until all sequences has calculated；

Step 4 eight judges that whether network error is less than anticipation error, if less than anticipation error, terminates training, otherwise judges whether iterations reaches maximum study number of times, jumps step 4 one, otherwise terminate training when less than maximum study number of times；

Step 5, input test data, it is determined that the secondary structure corresponding to protein sequence.

The present invention adopts BP neural network algorithm that secondary protein structure is predicted, and for the defect that BP neutral net exists, the learning process of network is improved.In network learning procedure, adopt batch processing learning model, promote the parallel spatial of calculating process, improve network convergence speed；Secondly, it is optimized based on gradient method, introduces the learning rules of subsidiary momentum term and autoadapted learning rate, it is to avoid " oscillatory occurences " and be absorbed in local minimum.In the selection of network structure and correlation technique, six input coding modes and sliding window technique is adopted at input layer, hidden layer structure rule of thumb formula and sliding window size are configured, and the classification of secondary protein structure is carried out output prediction according to DSSP algorithm by output layer.

Accompanying drawing explanation

Fig. 1 is sliding window schematic diagram, and window size is 2n+1, and window both sides length is n, and amino acid residue to be predicted is the residue I of window center position, and the contiguous amino acid residues of input network is YFQSMSVKGR and YSILKQIG, then the desired output of network is then T.Move to right a position with rear hatch, the structure of next residue is predicted, then amino acid residue to be predicted becomes Y, it is desirable to be output as E.

Fig. 2 is six coded system exemplary plot, the alanine (A) of window center be encoded to 000011, on the right side of window, isoleucine (I) is encoded to 010000.125.

Fig. 3 is the BP neural network model for secondary protein structure, and input layer has N number of neuron, and hidden layer has T neuron, and output layer has M neuron.

Fig. 4 is BP neural metwork training flow chart, is that training sample set is carried out repetition learning, makes the process that network is restrained gradually.

Detailed description of the invention

A kind of secondary protein structure prediction method based on BP neutral net, adopts following method to be embodied as:

Detailed description of the invention one: illustrate present embodiment below in conjunction with Fig. 1, Fig. 2, Fig. 3.

Step one, choose from PDB one group of α spiral, β-pleated sheet and curling three class formations account for normal rates protein structural database composition training sample set, three class formation ratios are all between 20%～40%, and coiled structure is on the high side；

Step 2 one, adopt six input coding protein amino acid sequence information.Wherein using five dimension two-stage system vector representation amino acid classes, remaining one is used for representing the residue side-play amount relative to window center position.Distance center position is more near, and the impact of residues Structures to be predicted is more big, and this value is also more big；Distance is more remote, then this value is more little.Assume that relative to the side-play amount at center be n, then this value is 2^-n, as shown in Figure 1.

When before sequence is started, several residues carry out structure prediction, the portion in window does not have residue information, here it is the null value problem that can run in an encoding process.In this case, the mode adopted herein is by this position zero, for instance when adopt six coded systems, null value position be encoded to 000000.By by null value position zero so that it is as a kind of incoming network of particular value, reduce the null value impact on e-learning.

Step 2 two, the selected input range of employing sliding window, sliding window technique is by utilizing a window to move on protein sequence to be predicted, the input data of input layer being carried out scope selection.The center of window is amino acid residue to be predicted, the positional symmetry of window both sides, and length is identical, comprises the residue information adjacent with position to be predicted.When carrying out residues Structures prediction, all amino acid residue information in window being inputted network, the output of network output layer is then the secondary structure of window center position amino acid residue, as shown in Figure 2.

Window is excessive or too small all predictablity rate and network training efficiency can be produced impact, after deliberation and test, selects 17 as sliding window size.

Step 3, BP Neural Network Training Parameter initialize, sliding window position initialization.In Fig. 3, network parameter is arranged as follows:

Sample input vector: I_a=(i₁, i, i₃..., i_n), a=1,2,3 ..., Q

Input layer input vector: X_a=(x₁, x₂,x₃..., x_n), a=1,2,3 ..., Q

Hidden layer output vector: H_a=(h₁,h₂,h₃..., h_t), a=1,2,3 ..., Q

The actual output vector of output layer: O_a=(o₁, o₂,o₃..., o_m), a=1,2,3 ..., Q

Output layer desired output vector: P_a=(p₁, p₂, p₃..., p_m), a=1,2,3 ..., Q

Input layer is to the connection weights of hidden layer: ω_ij, represent input layer x_iTo hidden layer h_jConnection weights, i=1,2,3 ..., N, j=1,2,3 ..., T

Hidden layer is to the connection weights of output layer: ε_jk, represent hidden layer h_jO to output layer_kConnect weights, j=1,2,3 ..., T, k=1,2,3

The threshold value of hidden layer neuron: θ_j, represent hidden layer neuron h_jCorresponding threshold value, j=1,2,3 ..., T

The neuronic threshold value of output layer: μ_k, represent output layer neuron o_kCorresponding threshold value, k=1,2,3

The learning rate of weighed value adjusting: δ

Excitation function selects Sigmoid function:

Step 3 one, BP network parameter initialize.Network is connected weights ω_ij、ε_jkAnd threshold θ_j、μ_kCarry out random assignment.Owing to adopting Sigmoid function as transmission function, function will input the input of the little scope of boil down on a large scale, but when input becomes larger, the change of Sigmoid just levels off to 0, it is only small that the now produced impact of the change of parameter becomes, the assignment interval selection of random assignment is between (-1,1)；

Step 3 two, sliding window position initialization, align first residue by sliding window；

BP neutral net instructs flow process as shown in Figure 4:

The output of step 4 one, hidden layer calculates, and chooses training sample input network, it is thus achieved that network input and desired output, calculates hidden layer output.Concentrate from training sample and choose training sample I_a, and by I_aIt is converted into network input X_aWith desired output P_a；Utilize input layer input X_a, hidden layer connect weights ω_ijAnd threshold θ_jCalculate hidden layer intermediate value, obtain the output h of hidden layer neuron through excitation function conversion_j:

$h_i=f\left(\underset{i=1}{\overset{N}{\mathrm{\Σ}}}{x}_i{\mathrm{\ω}}_{ij}-{\mathrm{\θ}}_j\right)$

The output of step 4 two, output layer calculates, and utilizes each neuron output h of hidden layer_j, output layer connect weights ε_jkAnd threshold value μ_kDraw output layer intermediate value, draw the actual output o of network through excitation function conversion_k:

$o_k=f(\underset{j=1}{\overset{T}{\Σ}}{h}_j{\mathrm{\ϵ}}_{jk}-{\mathrm{\μ}}_k)$

Step 4 three, output layer weights Error Calculation, calculate each neuronic correction error of output layer.Comparing cell actual output o_kWith desired output p_kBetween difference, based on minimum mean square error criterion, calculate each neuronic correction error of output layerAnd export h according to hidden layer_jWith correction errorIt is adjusted, it is thus achieved that new output layer neuron weights ε '_jkWith threshold value μ '_k:

$d_k^a=o_k(p_k-o_k)(1-o_k)$

${\mathrm{\ϵ}}_{jk}^{\′}={\mathrm{\ϵ}}_{jk}+\mathrm{\δ}*d_k^a*h_j$

${\mathrm{\μ}}_k^{\′}={\mathrm{\μ}}_k+\mathrm{\δ}*d_k^a$

Wherein,Represent and a sample is inputted, output layer neuron o_kCorrection error.

Step 4 four, hidden layer weights Error Calculation, comparing cell actual output o_kWith desired output p_kBetween difference, utilize output layer correction errorCalculate each neuronic correction error of output layerAnd export h based on hidden layer_j, output layer connect weights ε_jkX is inputted with input layer_iWeights are adjusted, it is thus achieved that new hidden layer neuron weights ω '_ijAnd threshold θ_j:

$e_j^a=h_j\left(1-h_j\right)\underset{k=1}{\overset{M}{\mathrm{\Σ}}}{d}_k^a{\mathrm{\ϵ}}_{jk}$

${\mathrm{\ω}}_{ij}^{\′}={\mathrm{\ω}}_{ij}+\mathrm{\δ}*e_j^a*x_i$

${\mathrm{\θ}}_i^{\′}={\mathrm{\θ}}_j+\mathrm{\δ}*e_j^a$

Wherein,Represent and a sample is inputted, output layer neuron h_jCorrection error.

Neuronic weights and threshold value, according to Error Calculation weighed value adjusting amount, are adjusted by each neuronic weights in step 4 five, reversely adjustment network；

Step 4 six, sliding window move to right one, skip to 41, until whole protein sequence has calculated；

Step 4 seven, select next training sample to learn, skip to 41, until all sequences has calculated；

Step 4 eight judges that whether network error is less than anticipation error, if less than anticipation error, terminates training, otherwise judges whether iterations reaches maximum study number of times, jumps step 4 one, otherwise terminate training when less than maximum study number of times；

Step 5, input protein data to be tested, it is determined that the secondary structure corresponding to protein sequence.

According to hydrogen bonding pattern, know together other eight kinds of secondary structure conformation, identification feature according to each of which and architectural feature, these structure conformation are divided into three major types, including: (1) helical structure, including 310 spirals (G), alpha-helix core (H), three kinds of spirals of π-spiral (I)；(2) foldable structure, including the hydrogen bond in two kinds of beta-pleated sheets to type, in parallel and antiparallel bridge, independent bridge construction is designated B, i.e. β-bridge, and the pleated sheet containing β-projection is designated E；(3) circulus, including all the other all ordinary constructions (T, S, L etc.).

Utilizing neutral net to carry out protein secondary structure prediction, the actual tendency parameter being output as the corresponding three kinds of structures of this residue of wherein network, tendency parameter corresponding to structure is more big, then illustrate that the probability that residue is this kind of structure is more high.Assume that the actual of network is output as a, b, c, then when a is maximum, residue is helical structure, is foldable structure when b is maximum, is circulus when c is maximum.

The overall accuracy rate evaluation methodology adopted:

Wherein, N is the sum of amino acid residue to be predicted, C_allFor overall accuracy rate, C_Spiral、C_FoldingAnd C_Ring-typeRepresent the number of residues of correctly predicted spiral, folding and circulus respectively.

Claims (15)

1. one kind based on BP neutral net secondary protein structure prediction method, it is characterised in that:

Adopt following step:

Step one, choose from PDB one group of α spiral, β-pleated sheet and curling three class formations account for normal rates protein structural database composition training sample set, ratio is between 20%～40%, and coiled structure is slightly many；

Step 2, adopt the amino acid sequence information of six coded proteins, as the input of BP neutral net, adopt the selected input range of sliding window；

Step 3, BP Neural Network Training Parameter initialize, sliding window position initialization；

Step 4, setting maximum iterations and minimum expectation error, the one by one sample in input sample set, training produces the BP neutral net for predicting；

In step 4, training produces the following steps of BP neutral net and realizes:

Step 4 one, from training set of stereotypes choose training sample, be converted into network input, calculate hidden layer output；

Step 4 two, utilize hidden layer calculate obtain output, calculate output layer output；

The actual output of step 4 three, comparing cell and desired output, calculate output layer weights error；

Step 4 four, utilize output layer correction error and hidden layer output, calculate hidden layer weights error；

Step 4 five, according to Error Calculation weighed value adjusting amount, reversely adjust hidden layer and the neuronic weights of output layer and threshold value in network；

Step 4 six, sliding window move to right one, skip to 41, until whole protein sequence has calculated；

Step 4 seven, select next training sample to learn, skip to 41, until all sequences has calculated；

Step 4 eight judges that whether network error is less than anticipation error, if less than anticipation error, terminates training, otherwise judges whether iterations reaches maximum study number of times, jumps step 4 one, otherwise terminate training when less than maximum study number of times；

Step 5, input protein data to be tested, it is determined that the secondary structure corresponding to protein sequence.

2. according to claim 1 based on BP neutral net secondary protein structure prediction method, it is characterized in that:

Adopting the amino acid sequence information of six input coding protein in step 2, wherein five represent amino acid classes, and one is used for representing the residue side-play amount relative to window center position, it is assumed that be n relative to the side-play amount at center, then this value is 2^-n。

3. according to claim 1 based on BP neutral net secondary protein structure prediction method, it is characterized in that:

Adopting the selected input range of sliding window in step 2, the center of window is amino acid residue to be predicted, the positional symmetry of window both sides, and length is identical, comprises the residue information adjacent with position to be predicted.

4. according to claim 1 based on BP neutral net secondary protein structure prediction method, it is characterized in that:

In step 3, initialized network connects weights ω_ij、ε_jkAnd threshold θ_j、μ_kFor random assignment, the assignment interval selection of random assignment is between (-1,1).

5. according to claim 1 based on BP neutral net secondary protein structure prediction method, it is characterized in that:

Step 4 adopts the transmission function that Sigmoid function calculates as neuron.

6. according to claim 1 and 5 based on BP neutral net secondary protein structure prediction method, it is characterized in that:

Step 4 one utilizes formula one to calculate hidden layer output, wherein ω_ijAnd θ_jHidden layer connects weights and threshold value, x_iInput for input layer i-th:

Formula one, $h_j=f(\underset{i=1}{\overset{N}{\Σ}}{x}_i{\mathrm{\ω}}_{ij}-{\mathrm{\θ}}_j).$

7. according to claim 1 and 5 based on BP neutral net secondary protein structure prediction method, it is characterized in that:

Step 4 two utilizes formula two to calculate output layer output, wherein, and h_jExport for hidden layer, ε_jkAnd μ_kWeights and threshold value is connected for output layer:

Formula two, $o_k=f(\underset{j=1}{\overset{T}{\Σ}}{h}_j{\mathrm{\ϵ}}_{jk}-{\mathrm{\μ}}_k).$

8. according to claim 1 based on BP neutral net secondary protein structure prediction method, it is characterized in that:

Step 4 three, based on minimum mean square error criterion, utilizes formula four to calculate each neuronic correction error of output layerAnd export h according to hidden layer_jWith correction errorFormula four and five is utilized to calculate new output layer neuron weights ε '_jkWith threshold value μ '_k,

Formula three $d_k^a=o_k(p_k-o_k)(1-o_k)$

Formula four ${\mathrm{\ϵ}}_{jk}^{\′}={\mathrm{\ϵ}}_{jk}+\mathrm{\δ}*d_k^a*h_j$

Formula five ${\mathrm{\μ}}_k^{\′}={\mathrm{\μ}}_k+\mathrm{\δ}*d_k^a.$

9. according to claim 1 based on BP neutral net secondary protein structure prediction method, it is characterized in that:

Step 4 four comparing cell actual output o_kWith desired output p_kBetween difference, utilize output layer correction errorFormula six is utilized to calculate each neuronic correction error of output layerH is exported based on hidden layer_j, output layer connect weights ε_jkX is inputted with input layer_i, utilize formula seven and eight to calculate new hidden layer neuron weights ω '_ijWith threshold θ '_j:

Formula six $e_j^a=h_j(1-h_j)\underset{k=1}{\overset{M}{\Σ}}{d}_k^a{\mathrm{\ϵ}}_{jk}$

Formula seven ${\mathrm{\ω}}_{ij}^{\′}={\mathrm{\ω}}_{ij}+\mathrm{\δ}*e_j^a*x_i$

Formula eight ${\mathrm{\θ}}_j^{\′}={\mathrm{\θ}}_j+\mathrm{\δ}*e_j^a.$

10. according to claim 1 based on BP neutral net secondary protein structure prediction method, it is characterized in that:

Step 4 five introduces the learning rules with momentum term, and weighed value adjusting quantitative change turns to:

${\mathrm{\Δ\ω}}_{ij}\left(t\right)=\mathrm{\δ}(1-\mathrm{\α})e_j^a{x}_i+\mathrm{\α}*{\mathrm{\Δ\ω}}_{ij}(t-1)$

Wherein, α is momentum term coefficient, and value is between (0,1).

11. according to claim 1 based on BP neutral net secondary protein structure prediction method, it is characterized in that:

Whether step 4 eight looks first at network error less than anticipation error, if less than anticipation error, terminating training, otherwise continuing iteration until reaching maximum study number of times.

12. according to claim 1 based on BP neutral net secondary protein structure prediction method, it is characterized in that:

Step 5 is in input test data, it is determined that adopt overall accuracy rate evaluation methodology during secondary structure corresponding to protein sequence:

Wherein, N is the sum of amino acid residue to be predicted, C_allFor overall accuracy rate, C_Spiral、C_FoldingAnd C_Ring-typeRepresent the number of residues of correctly predicted spiral, folding and circulus respectively.

13. protein coding method according to claim 2, null value position be encoded to 000000, as a kind of incoming network of particular value, reduce the null value impact on e-learning.

14. sliding window establishing method according to claim 3, time initial, in window, position is first residue, n null value position of left side polishing；Mend a null value position on the right side of during window alignment n residue of inverse, by that analogy, on the right side of when window is directed at last residue, mend n null value position.

15. according to claim 1,8 and 9 based on BP neutral net secondary protein structure prediction method, it is characterized in that:

Situation of change according to network error, dynamically arranges learning rate δ, makes network ensureing while pace of learning, takes into account the stability of network, if △ E represent the total error of twice study in front and back difference, introducing parameter beta₁And β₂Carrying out regularized learning algorithm rate, learning rate change formula is:

$\mathrm{\&delta;}=\left\{\begin{array}{cc}\frac{{\mathrm{\&beta;}}_1}{k}\mathrm{\&delta;}& (\mathrm{\&Delta;}E>0)\\ \mathrm{\&delta;}& (\mathrm{\&Delta;}E=0)\\ {\mathrm{k\&beta;}}_2\mathrm{\&delta;}& (\mathrm{\&Delta;}E<0)\end{array}\right.$

Wherein, β₁And β₂As Dynamic gene, β₁< 1, β₂> 1, k is read-around ratio.