CN103390119B - A kind of Binding site for transcription factor recognition methods - Google Patents

Abstract

The invention belongs to molecular biosciences infomation detection field, be specifically related to a kind of based on condition random field technology, merge ChIP-chip microarray data and ChIP-seq? the Binding site for transcription factor recognition methods of DNA sequencing data.The present invention includes: set up the condition random field models; Obtain ChIP-chip laboratory values, identify corresponding states value; Obtain ChIP-seq laboratory values and identify corresponding states value; The accuracy of identification of test condition random field models; N-th DNA base fragment identification probability of Weighted Fusion recognition result; Relatively with identify Binding site for transcription factor.The present invention utilizes condition random field to merge the detection data identification Binding site for transcription factor of ChIP-chip and ChIP-seq experiment.Verify by experiment, in recognition accuracy, the inventive method is higher than the recognition methods adopting monotechnics.

Description

A kind of Binding site for transcription factor recognition methods

Technical field

The invention belongs to molecular biosciences infomation detection field, be specifically related to a kind of based on condition random field technology, merge the Binding site for transcription factor recognition methods of ChIP-chip microarray data and ChIP-seqDNA sequencing data.

Background technology

The functional protein that transcription factor is transcribed as controlling gene, the identification of its binding site on DNA gene promoter region plays great role for the research of gene transcription regulation mechanism, and it is the Focal point and difficult point in life science.Along with the progress of information science experimental technique, also make rapid progress for the laboratory facilities obtaining molecular biosciences information.Wherein, the standard test detecting region of DNA territory specified protein bonding state within the scope of full-length genome is chromatin imrnunoprecipitation reaction (Chromatinimmunoprecipitaion, ChIP), chromatin imrnunoprecipitation reaction (ChIP) ChIP-chip and the ChIP-seq technology produced that combines with genetic chip or DNA sequencing two kinds of detection techniques is respectively then detect two large major techniques of chromatin imrnunoprecipitation reaction experiment result.

Due to the difference of Cleaning Principle, the experimental data that ChIP-chip and ChIP-seq experimental technique produces also exists respective relative merits.Genetic chip directly measures biological information, and its susceptibility is higher, but compared to high flux DNA sequencing, its measuring accuracy is not high, and test specificity is lower; High flux DNA sequencing then has higher detection resolution, and its measuring accuracy is higher, and test specificity is better, but owing to being carry out indirect inspection to biological information, its susceptibility is lower.Visible, although the experimental data obtained respectively by the experimental approach that ChIP-chip with ChIP-seq is different has relative merits different separately, but have obvious complementarity, therefore, the detection data merging these two kinds of technology acquisitions can obtain more complete Detection Information.

As the pattern recognition problem for one-dimensional sequence, analyze the binding site whether current dna base fragment is specific function albumen, not only to consider whether the biochemical characteristic of current base fragment is applicable to specific function protein combination, also needs to consider that the biochemical characteristic of the multiple base fragment in its both sides is on the impact of specific function protein binding capacity.Specific to the identification problem of Binding site for transcription factor, should consider in identifying exactly simultaneously binding site base fragment to be identified and near the characteristic of multiple base fragment.Numerous in the mode identification technology of one-dimensional sequence, due to condition random field (ConditionalRandomFields, CRF) technology has the directionless constraint of feature selecting, and feature locations supports the characteristics such as long-range correlation, makes it be applicable to very much the identification of Binding site for transcription factor.

Summary of the invention

The object of this invention is to provide a kind of more high-precision based on the fusion chip data of condition random field and the Binding site for transcription factor recognition methods of DNA sequencing data.

The object of the present invention is achieved like this:

(1) set up the condition random field models:

$p\left(y\right|x)=\frac{1}{Z\left(x\right)}\exp (\underset{i,k}{\mathrm{\Σ}}{\mathrm{\λ}}_k{t}_k(y_{i-1},y_i,x,i)+\underset{i,l}{\mathrm{\Σ}}{\mathrm{\μ}}_l{s}_l(y_i,x,i))$

$Z\left(x\right)=\underset{y}{\mathrm{\Σ}}\exp (\underset{i,k}{\mathrm{\Σ}}{\mathrm{\λ}}_k{t}_k(y_{i-1},y_i,x,i)+\underset{i,l}{\mathrm{\Σ}}{\mathrm{\μ}}_l{s}_l(y_i,x,i))$

Wherein, x={x ₁, x ₂..., x _nrepresent the laboratory values of DNA base fragment; Y={y ₁, y ₂..., y _nthe corresponding states value of DNA base fragment, 1 represents it is Binding site for transcription factor, and 0 represents it is not Binding site for transcription factor; t _k(y _i-1, y _i, x, i) and represent that i-th base fragment is under current experiment detected value sequence x, state is y _ii-th base fragment and state be y _i-1the i-th-1 base fragment between transfer characteristic function; s _l(y _i, x, i) represent under current experiment detected value sequence x, the state of i-th base fragment is y _istatus flag function; λ _kand μ _lt respectively _k(y _i-1, y _i, x, i) and s _l(y _i, x, i) and corresponding weights, represent the importance of each fundamental function; Z (x) is standardizing factor, makes p (y|x) be positioned between [0,1];

(2) the ChIP-chip laboratory values x={x of DNA base fragment is obtained ₁, x ₂..., x _n, according to conditional random field models, identify the state value y={y of corresponding DNA base fragment ₁, y ₂..., y _n;

(3) the ChIP-seq laboratory values x={x of DNA base fragment is obtained ₁, x ₂..., x _n, according to conditional random field models, identify the state value y={y of corresponding DNA base fragment ₁, y ₂..., y _n;

(3) accuracy of identification of test condition random field models:

$S_n=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}},$

$S_P=\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}},$

$A_c=\frac{S_n+S_p}{2},$

Wherein, susceptibility S _n, specificity S _p, accuracy rate A _c, TP represents the predicted correct number of Binding site for transcription factor; FN represents the number of the predicted mistake of Binding site for transcription factor; TN represents the predicted correct number of non-transcribed factor binding site; FP represents the number of the predicted mistake of non-transcribed factor binding site;

(4) to the n-th DNA base fragment, the probability being identified as Binding site for transcription factor by ChIP-chip laboratory values is used represent, the probability being identified as non-transcribed factor binding site is used represent, the probability being identified as Binding site for transcription factor by ChIP-seq laboratory values is used represent, the probability being identified as non-transcribed factor binding site is used represent, the n-th DNA base fragment identification probability of Weighted Fusion recognition result is expressed as:

$p_1^{\left(n\right)}=p_{\mathrm{lchip}}^{\left(n\right)}\×w_1+p_{1\mathrm{seq}}^{\left(n\right)}\×w_2,$

$p_0^{\left(n\right)}=p_{0\mathrm{chip}}^{\left(n\right)}\×w_1+p_{0\mathrm{seq}}^{\left(n\right)}\×w_2,$

Blending weight w1 and w2 is the recognition accuracy with recognition result, w ₁+ w ₂=1;

(5) compare with then this base fragment is identified as Binding site for transcription factor; then this base fragment is identified as the non-binding site of transcription factor.

Beneficial effect of the present invention is:

The present invention utilizes condition random field to merge the detection data identification Binding site for transcription factor of ChIP-chip and ChIP-seq experiment.Verify by experiment, in recognition accuracy, the inventive method is higher than the recognition methods adopting monotechnics.

Accompanying drawing explanation

Fig. 1 is Binding site for transcription factor recognition methods process flow diagram of the present invention;

Fig. 2 is the training process flow diagram of conditional random field models.

Embodiment

Below in conjunction with accompanying drawing, the present invention is described further:

Based on the fusion chip data of condition random field and the Transcription Factor Binding Sites Prediction method of DNA sequencing data, comprise the following steps:

(1) adopt condition random field recognition technology to adopt ChIP-chip microarray data, Binding site for transcription factor is tentatively identified;

(2) adopt condition random field recognition technology to adopt ChIP-seqDNA sequencing data, Binding site for transcription factor is tentatively identified;

(3) adopt the mode of decision level fusion, Binding site for transcription factor is finally identified.Namely to the base fragment of same position, Weighted Fusion respectively from the preliminary recognition result of Binding site for transcription factor of ChIP-chip microarray data and ChIP-seqDNA sequencing data, and draws final recognition result.Wherein, blending weight is directly proportional to the recognition accuracy of preliminary recognition result, and two blending weights and be 1.

The representation of the condition random field that this method adopts is as follows:

$p\left(y\right|x)=\frac{1}{Z\left(x\right)}\exp (\underset{i,k}{\mathrm{\Σ}}{\mathrm{\λ}}_k{t}_k(y_{i-1},y_i,x,i)+\underset{i,l}{\mathrm{\Σ}}{\mathrm{\μ}}_l{s}_l(y_i,x,i))$

$Z\left(x\right)=\underset{y}{\mathrm{\Σ}}\exp (\underset{i,k}{\mathrm{\Σ}}{\mathrm{\λ}}_k{t}_k(y_{i-1},y_i,x,i)+\underset{i,l}{\mathrm{\Σ}}{\mathrm{\μ}}_l{s}_l(y_i,x,i))$

Wherein, x={x ₁, x ₂..., x _nrepresent the laboratory values of DNA base fragment; Y={y ₁, y ₂..., y _nthe corresponding states value of DNA base fragment, only have two states value here, 1 represents it is Binding site for transcription factor, and 0 represents it is not Binding site for transcription factor; t _k(y _i-1, y _i, x, i) and represent that i-th base fragment is under current experiment detected value sequence x, state is y _ii-th base fragment and state be y _i-1the i-th-1 base fragment between transfer characteristic function; s _l(y _i, x, i) represent under current experiment detected value sequence x, the state of i-th base fragment is y _istatus flag function; λ _kand μ _lt respectively _k(y _i-1, y _i, x, i) and s _l(y _i, x, i) and corresponding weights, represent the importance of each fundamental function; Z (x) is standardizing factor, and p (y|x) is positioned between [0,1].Condition random field fundamental function selected by the present invention comprises status flag function and transfer characteristic function.Wherein, status flag selection window is 1, namely in conjunction with the detected value constitutive characteristic of each base fragment of current base fragment and front and back.It is 1 that transfer characteristic is simply taken as adjacent segment characterizations, otherwise is 0.Fundamental function then elect as each feature self, quadratic sum evolution, finally decided the importance of each fundamental function by weight.Adopt above-mentioned model, when obtaining the laboratory values x={x of DNA base fragment ₁, x ₂..., x _nafter, the state value y={y of corresponding DNA base fragment can be identified ₁, y ₂..., y _n, for judging whether each base fragment is Binding site for transcription factor.

The inventive method embodiment is as shown in Figure 1, specific as follows:

1. the preliminary recognition result of base fragment ChIP-chip data CRF model Binding site for transcription factor

The condition random field CRF modelling that base fragment ChIP-chip data transcription factor binding site tentatively identifies is as follows:

$p\left(y\right|x)=\frac{1}{Z\left(x\right)}\exp (\underset{i,k}{\mathrm{\Σ}}{\mathrm{\λ}}_k{t}_k(y_{i-1},y_i,x,i)+\underset{i,l}{\mathrm{\Σ}}{\mathrm{\μ}}_l{s}_l(y_i,x,i))$

$Z\left(x\right)=\underset{y}{\mathrm{\Σ}}\exp (\underset{i,k}{\mathrm{\Σ}}{\mathrm{\λ}}_k{t}_k(y_{i-1},y_i,x,i)+\underset{i,l}{\mathrm{\Σ}}{\mathrm{\μ}}_l{s}_l(y_i,x,i))$

Wherein, x={x ₁, x ₂..., x _nrepresent the ChIP-chip laboratory values of DNA base fragment; Y={y ₁, y ₂..., y _nthe corresponding states value of DNA base fragment, only have two states value here, 1 represents it is Binding site for transcription factor, and 0 represents it is not Binding site for transcription factor; t _k(y _i-1, y _i, x, i) and represent that i-th base fragment is under current experiment detected value sequence x, state is y _ii-th base fragment and state be y _i-1the i-th-1 base fragment between transfer characteristic function; s _l(y _i, x, i) represent under current experiment detected value sequence x, the state of i-th base fragment is y _istatus flag function; λ _kand μ _lt respectively _k(y _i-1, y _i, x, i) and s _l(y _i, x, i) and corresponding weights, represent the importance of each fundamental function; Z (x) is standardizing factor, and p (y|x) is positioned between [0,1].Condition random field fundamental function selected by the present invention comprises status flag function and transfer characteristic function.Wherein, status flag selection window is 1, namely in conjunction with the detected value constitutive characteristic of each base fragment of current base fragment and front and back.It is 1 that transfer characteristic is simply taken as adjacent segment characterizations, otherwise is 0.Fundamental function then elect as each characteristic self, quadratic sum evolution, finally decided the importance of each fundamental function by weight.

Adopt above-mentioned model, when obtaining the ChIP-chip laboratory values x={x of DNA base fragment ₁, x ₂..., x _nafter, the state value y={y of corresponding DNA base fragment can be identified ₁, y ₂..., y _n, as the preliminary recognition result of each base fragment Binding site for transcription factor.

2. the preliminary recognition result of base fragment ChIP-seq data CRF model Binding site for transcription factor

The condition random field CRF modelling that base fragment ChIP-seq data transcription factor binding site tentatively identifies is as follows:

$p\left(y\right|x)=\frac{1}{Z\left(x\right)}\exp (\underset{i,k}{\mathrm{\Σ}}{\mathrm{\λ}}_k{t}_k(y_{i-1},y_i,x,i)+\underset{i,l}{\mathrm{\Σ}}{\mathrm{\μ}}_l{s}_l(y_i,x,i))$

$Z\left(x\right)=\underset{y}{\mathrm{\Σ}}\exp (\underset{i,k}{\mathrm{\Σ}}{\mathrm{\λ}}_k{t}_k(y_{i-1},y_i,x,i)+\underset{i,l}{\mathrm{\Σ}}{\mathrm{\μ}}_l{s}_l(y_i,x,i))$

Wherein, x={x ₁, x ₂..., x _nrepresent the ChIP-seq laboratory values of DNA base fragment; Y={y ₁, y ₂..., y _nthe corresponding states value of DNA base fragment, only have two states value here, 1 represents it is Binding site for transcription factor, and 0 represents it is not Binding site for transcription factor; t _k(y _i-1, y _i, x, i) and represent that i-th base fragment is under current experiment detected value sequence x, state is y _ii-th base fragment and state be y _i-1the i-th-1 base fragment between transfer characteristic function; s _l(y _i, x, i) represent under current experiment detected value sequence x, the state of i-th base fragment is y _istatus flag function; λ _kand μ _lt respectively _k(y _i-1, y _i, x, i) and s _l(y _i, x, i) and corresponding weights, represent the importance of each fundamental function; Z (x) is standardizing factor, and p (y|x) is positioned between [0,1].Condition random field fundamental function selected by the present invention comprises status flag function and transfer characteristic function.Wherein, status flag selection window is 1, namely in conjunction with the detected value constitutive characteristic of each base fragment of current base fragment and front and back.It is 1 that transfer characteristic is simply taken as adjacent segment characterizations, otherwise is 0.Fundamental function then elect as each characteristic self, quadratic sum evolution, finally decided the importance of each fundamental function by weight.

Adopt above-mentioned model, when obtaining the ChIP-seq laboratory values x={x of DNA base fragment ₁, x ₂..., x _nafter, the state value y={y of corresponding DNA base fragment can be identified ₁, y ₂..., y _n, as the preliminary recognition result of each base fragment Binding site for transcription factor.

3. the fusion recognition result of Binding site for transcription factor

The present invention adopts the mode of decision level fusion to obtain final Binding site for transcription factor recognition result.Namely to the base fragment of same position, the mode of employing weighting merges the preliminary recognition result of base fragment CRF model Binding site for transcription factor respectively from ChIP-chip microarray data and ChIP-seqDNA sequencing data, and draws final Binding site for transcription factor recognition result.Be specially:

First, adopt the mode of cross validation, the accuracy of identification of condition random field training pattern is tested.Wherein, test index adopts susceptibility S _n(Sensitivity), specificity S _p(Specificity), accuracy rate A _c, they are defined as follows:

$S_n=\frac{\mathrm{TP}}{\mathrm{TP}+\mathrm{FN}}---\left(1\right)$

$S_P=\frac{\mathrm{TN}}{\mathrm{TN}+\mathrm{FP}}---\left(2\right)$

$A_c=\frac{S_n+S_p}{2}---\left(3\right)$

In formula, TP represents the predicted correct number of Binding site for transcription factor; FN represents the number of the predicted mistake of Binding site for transcription factor; TN represents the predicted correct number of non-transcribed factor binding site; FP represents the number of the predicted mistake of non-transcribed factor binding site.The accuracy rate A of condition random field training pattern _casking for of blending weight will be used for.

Secondly, for the n-th DNA base fragment, the probability being tentatively identified as Binding site for transcription factor based on ChIP-chip data is used represent, the probability being tentatively identified as non-transcribed factor binding site is used represent; The probability being tentatively identified as Binding site for transcription factor based on ChIP-seq data is used represent, the probability being tentatively identified as non-transcribed factor binding site is used represent.This base fragment identification probability then merging above-mentioned preliminary recognition result is expressed as:

$p_1^{\left(n\right)}=p_{\mathrm{lchip}}^{\left(n\right)}\×w_1+p_{1\mathrm{seq}}^{\left(n\right)}\×w_2---\left(4\right)$

$p_0^{\left(n\right)}=p_{0\mathrm{chip}}^{\left(n\right)}\×w_1+p_{0\mathrm{seq}}^{\left(n\right)}\×w_2,$

In formula, blending weight w ₁and w ₂be designed to the recognition accuracy with preliminary recognition result, i.e. the accuracy rate A of condition random field training pattern _cbe directly proportional, and two blending weights and be 1.

Finally, compare with if then this base fragment is identified as Binding site for transcription factor; If then this base fragment is identified as the non-binding site of transcription factor.

4. experimental verification

4.1 data prediction

4.1.1ChIP-chip data prediction

The ChIP-chip microarray data that this experiment adopts derives from international large-scale public biomolecule information database GEO database (GSE6892), the bonding state of transcription factor STAT1 on this Data Detection human body 44 genes.Each data measured is that 50bp is long, but owing to have employed tile test format, exists and partly overlap between adjacent two data measureds.We again split data and combine, the base fragment be connected of the final 38bp of acquisition, and calculate the ChIP-chip detected value in each fragment.

4.1.2ChIP-seq data prediction

The ChIP-seq sequencing data that this experiment adopts derives from BCGSC database, is also the detection for human transcription factor STAT1 bonding state.When carrying out pretreated to the data of ChIP-seq, first raw data is converted to fastq form, then comparison is passed through, data measured is mapped in human genome, and calculate the data measured degree of covering of each base position on DNA, finally, the DNA base fragment that corresponding ChIP-chip data obtain, calculates the ChIP-seq data measured degree of covering value of each base fragment.

4.2 conditional random field models training

The fundamental function of conditional random field models comprises status flag function and transfer characteristic function.Wherein, status flag selection window is 1, namely in conjunction with the detected value constitutive characteristic of each base fragment of current base fragment and front and back.It is 1 that transfer characteristic is simply taken as adjacent segment characterizations, otherwise is 0.Fundamental function then elect as each feature self, quadratic sum evolution.Utilize STAT1 bonding state data on known DNA base fragment, train conditional random field models, training process flow diagram as shown in Figure 2.

The fusion of 4.3 conditional random field models recognition results

First, adopt the mode of 5 cross validations, the accuracy of identification of condition random field training pattern is tested.Test result is as shown in Table 1 and Table 2:

Table 1ChIP-chip data cross the result

Table 2ChIP-seq data cross the result

In table 1, the Average Accuracy recording the CRF model adopting the training of ChIP-chip data with five cross-validation methods is 57.68%., in table 2, the Average Accuracy recording the CRF model adopting the training of ChIP-seq data with five cross-validation methods is 63.76%.

Then, according to based on the CRF model training accuracy rate calculating blending weight of ChIP-chip data and ChIP-seq data being respectively: w1=0.475, w2=0.525.And to calculate the n-th DNA base fragment be accordingly the probability of Binding site for transcription factor with the probability for non-transcribed factor binding site

Finally, compare with if then the n-th DNA base fragment is identified as the binding site of transcription factor STAT1; If then the n-th DNA base fragment is identified as the non-binding site of transcription factor STAT1.Adopt the accuracy test of five cross-validation methods to fusion results as shown in table 3:

The cross validation results that table 3 merges

Visible, the Average Accuracy merging the recognition result that ChIP-chip data and ChIP-seq data obtain is 0.7697, significantly higher than the Average Accuracy being used alone the recognition result that ChIP-chip data and ChIP-seq data obtain.

Claims (1)

1. a Binding site for transcription factor recognition methods, is characterized in that:

(1) set up the condition random field models:

Wherein, x={x ₁, x ₂..., x _nrepresent the laboratory values of DNA base fragment; Y={y ₁, y ₂..., y _nthe corresponding states value of DNA base fragment, 1 represents it is Binding site for transcription factor, and 0 represents it is not Binding site for transcription factor; t _k(y _i-1, y _i, x, i) and represent that i-th base fragment is under current experiment detected value sequence x, state is y _ii-th base fragment and state be y _i-1the i-th-1 base fragment between transfer characteristic function; s _l(y _i, x, i) represent under current experiment detected value sequence x, the state of i-th base fragment is the status flag function of yi; λ _kand μ _lt respectively _k(y _i-1, y _i, x, i) and s _l(y _i, x, i) and corresponding weights, represent the importance of each fundamental function; Z (x) is standardizing factor, makes p (y|x) be positioned between [0,1];

(2) the ChIP-chip laboratory values x={x of DNA base fragment is obtained ₁, x ₂..., x _n, according to conditional random field models, identify the state value y={y of corresponding DNA base fragment ₁, y ₂..., y _n;

(3) the ChIP-seq laboratory values x={x of DNA base fragment is obtained ₁, x ₂..., x _n, according to conditional random field models, identify the state value y={y of corresponding DNA base fragment ₁, y ₂..., y _n;

(4) accuracy of identification of test condition random field models:

Wherein, susceptibility S _n, specificity S _p, accuracy rate A _c, TP represents the predicted correct number of Binding site for transcription factor; FN represents the number of the predicted mistake of Binding site for transcription factor; TN represents the predicted correct number of non-transcribed factor binding site; FP represents the number of the predicted mistake of non-transcribed factor binding site;

(5) to the n-th DNA base fragment, the probability being identified as Binding site for transcription factor by ChIP-chip laboratory values is used represent, the probability being identified as non-transcribed factor binding site is used represent, the probability being identified as Binding site for transcription factor by ChIP-seq laboratory values is used represent, the probability being identified as non-transcribed factor binding site is used represent, the n-th DNA base fragment identification probability of Weighted Fusion recognition result is expressed as:

Blending weight w ₁and w ₂for the recognition accuracy with recognition result, w ₁+ w ₂=1;

(6) compare with then this base fragment is identified as Binding site for transcription factor; then this base fragment is identified as the non-binding site of transcription factor.