Single-cell transcription factor prediction method based on deep learning and attention mechanism

Abstract

The application discloses a single-cell transcription factor prediction method based on deep learning and an attention mechanism, which comprises the following steps: obtaining single-cell chromatin accessibility analysis sequencing data, and preprocessing the data; then, data enhancement operation is performed to obtain enhanced sequencing data; regression peaks in the enhanced sequencing data are extracted as a feature vector S, forward and reverse enhanced sequencing data are spliced as a feature vector A, and DNA sequence data taken from a whole genome is converted into a feature vector U; the feature vector S, the feature vector A and the feature vector U are spliced, and then input into a deep network model to predict the probability of each transcription factor in a single cell, wherein the deep network model comprises a convolution module and a channel attention model.

Description

TECHNICAL FIELD

[0001] The present application relates to a transcription factor detection technology, in particular to a single-cell transcription factor prediction method based on deep learning and attention mechanism. BACKGROUND

[0002] Gene transcription regulation is an important mechanism for gene expression of organisms, which interacts with related receptors in organisms through transcription factors to ensure normal gene expression process of cells and can quickly respond to subtle changes in biological environment. At present, detecting transcription factor binding sites in the whole genome is an important part of exploring the mechanism of gene transcription regulation.

[0003] In recent years, thanks to the rapid development of high-throughput sequencing technology, single-cell sequencing technology has also entered a period of rapid development. From the perspective of single cells, we can more deeply and accurately understand important fields such as cancer heterogeneity, stem cell development and differentiation, human cell atlas, and gene transcription regulation. How to accurately and efficiently use the existing single-cell data to predict transcription factor binding sites is a very important and urgent problem to be solved.

[0004] At present, due to the progress of sequencing technology and the significant improvement of computing power, deep learning technology has been widely used in the prediction of transcription factor binding sites due to its advantages in data processing. The prediction model based on deep learning mainly includes convolutional neural network, recurrent neural network and hybrid neural network. At present, the prediction method based on convolutional neural network has relatively ideal effect. Zhou et al. proposed DeepSEA model, which can predict transcription factors by learning large-scale chromatin data information, and can also predict the influence of DNase I sensitivity and single nucleotide sensitivity on chromatin. Alipanahi et al. proposed DeepBind model, which uses convolutional neural network to realize the prediction of sequence specificity of DNA or RNA and protein binding.

[0005] Although the deep learning algorithm overcomes the shortcomings of the traditional machine learning method to some extent, there are still the following shortcomings: first, the training data used at present is relatively single, but other omics can also provide relevant feature information. Second, due to the deepening of the number of convolutional neural network layers and the black box characteristics of neural network, the explainability of the whole calculation model is also significantly weakened. SUMMARY

[0006] In view of the above shortcomings in the prior art, the single-cell transcription factor prediction method based on deep learning and attention mechanism provided by the present application solves the problem of low data feature extraction efficiency of the deep learning model.

[0007] In order to achieve the above-mentioned application purposes, the technical scheme adopted by the present application is:

[0008] A single-cell transcription factor prediction method based on deep learning and attention mechanism is provided, comprising the steps of:

[0009] Obtain single-cell chromatin accessibility analysis sequencing data and preprocess it, then perform data enhancement operation to obtain enhanced sequencing data;

[0010] Extract the regression peak in the enhanced sequencing data as a feature vector S, splice the forward and reverse enhanced sequencing data as a feature vector A, and convert the DNA sequence data taken from the whole genome into a feature vector U;

[0011] Splice the feature vector S, the feature vector A and the feature vector U, and input them into a deep network model to predict the probability of each transcription factor in the single cell, wherein the deep network model comprises a convolution module and a channel attention model.

[0012] The present application has the following advantages: by preprocessing the single-cell chromatin accessibility analysis sequencing data, the noise in the data can be reduced, and then the data is input into a deep learning model embedded with a channel attention module for recognition, thereby improving the speed and accuracy of feature extraction of the model for the required data.

[0013] Further, the input deep network model for predicting the probability of the transcription factor further comprises:

[0014] S31, determine whether the cell line to which the single-cell chromatin accessibility analysis sequencing data belongs is known, if yes, select the corresponding deep network model to predict, otherwise, go to step S32;

[0015] S32, input the spliced data into the deep network model corresponding to the cell lines GM12878, K562 and H1ESC respectively to predict the probability of the transcription factor;

[0016] S33, select the maximum value of the transcription factor prediction probability output by the three models as the transcription factor prediction probability of the single cell.

[0017] The above technical solution has the following advantages: the present application constructs corresponding deep network models for common cell lines, so that in the case of limited sequencing data information collected, the single cell can be quickly determined to be located in a specific cell line, and the probability of the transcription factor in the cell can be accurately obtained.

[0018] Further, the convolution module comprises a convolution layer, an activation layer, a pooling layer and a full connection layer connected in sequence; the channel attention model comprises a max pooling layer / average pooling layer, a shared multi-layer perception module, a full connection layer and a flattening layer connected in sequence; and the model structure of the deep network model is:

[0019] F1 = max_pooling(ReLU(conv1(S F )), F2 = max_pooling(ReLU(conv2(F1))

[0020] F3 = max_pooling(ReLU(conv3(F2)), F = ReLU(W1 F3)

[0021] Z i,n,k = sigmoid(W2 M c (F))

[0022] where S F is the spliced vector; ReLU(.) is the activation function; the convolution module includes a convolution layer, an activation layer, a pooling layer and a fully connected layer connected in turn; the channel attention model includes a max pooling layer / average pooling layer, a shared multi-layer perception module, a fully connected layer and a flattening layer connected in turn; F1, F2 and F3 are respectively the feature vector atlas of each convolution layer of the convolution module; W1 and W2 are respectively the weight matrix in the two fully connected layers; and are respectively the features calculated by the global average pooling and the global maximum pooling in the channel attention model; sigma is the sigmoid activation function; W0 and W3 are respectively two layers of parameters in the multi-layer perception model; M c (F) is the output of the channel attention model; Z i,n,k is the coincidence probability of the transcription factor k and the nth cell peak in each cell i; k e 1…M, M is the total number of transcription factors in each cell in the deep network model, and n e 1…N, N is the number of cell peaks that can be read in each cell.

[0023] The beneficial effects of the above technical solutions are that the convolution model with a small number of layers and the attention mechanism module are combined, which improves the efficiency of the model in extracting features and provides the model with interpretability.

[0024] Further, the training method of the deep network model comprises:

[0025] Pretreatment of DNA sequence data: select the DNA sequence data of the whole genome, and cut it into 200bp segments, with a sliding cutting interval of 50bp between each segment; obtain the reverse strand of each segment, and expand each reverse strand to a 1000bp segment with 200bp as the center, and then convert it into Mapability Data;

[0026] Preprocessing batch organization chromatin accessibility analysis sequencing data: obtain batch organization chromatin accessibility analysis sequencing data in cell lines GM12878, K562, H1ESC, and perform shearing operation thereon, then map the sheared mutual data to human genome hg19 using Bowtie2 tool, and perform cutting using samtools and Picard; convert the cut bam file to bigwig file using deepTools2, and then splice all bigwig files into a data matrix using bigWigMerge tool;

[0027] Encode the DNA sequence data before preprocessing into a 4*1000 feature vector S1 using one-hot; splice the preprocessed batch organization chromatin accessibility analysis sequencing data corresponding to each cell line in the forward and reverse directions as feature vectors A1, A2 and A3; convert the Mapability Data into a 2*1000 feature vector U1;

[0028] Splice feature vector S1, feature vector A1 and feature vector U1, feature vector S1, feature vector A2 and feature vector U1, and feature vector S1, feature vector A3 and feature vector U1, respectively, and input the spliced three vectors into three deep neural networks for training to obtain the deep network model corresponding to cell lines GM12878, K562 and H1ESC.

[0029] The beneficial effects of the above technical solutions are: the present scheme uses batch organization chromatin accessibility analysis sequencing data, DNA sequence data and Mapability Feature data for pre-training, which can fully use the data characteristics of each data, thereby ensuring the accuracy of transcription factor prediction in single-cell chromatin accessibility analysis sequencing data.

[0030] Further, the method for preprocessing single-cell chromatin accessibility analysis sequencing data comprises:

[0031] The peak read information of the single-cell chromatin accessibility analysis sequencing data is screened using the ENCODE single-cell chromatin accessibility analysis sequencing data processing method.

[0032] The bam file of the peak read information is converted to a bigwig file using deepTools2, and then all bigwig files are spliced into a data matrix using bigWigMerge tool.

[0033] The beneficial effects of the above technical solutions are that preprocessing single-cell chromatin sequencing data using the ENCODE process can maximize the preservation of original signal information.

[0034] Further, the data enhancement operation comprises:

[0035] The cisTopic method is used to calculate the potential characteristics of single cells, and the cosine similarity data index is used to calculate the similarity score between cells.

[0036] The 100 most similar adjacent cells are selected as the data for single-cell chromatin accessibility analysis sequencing data expansion, and the single cells and the selected adjacent cells are collected as enhanced sequencing data.

[0037] The beneficial effects of the above technical solutions are: the above-mentioned scheme can increase the coverage of chromatin accessibility while maintaining cell specificity, and such operation can effectively reduce the batch effect.

[0038] Further, the single-cell transcription factor prediction method further comprises calculating the activity factor score of the transcription factor according to the probability of the transcription factor:

[0039]

[0040]

[0041]

[0042] wherein, are the coincidence probabilities of the transcription factor k and the nth cell peak in the cell i output by the final network model, and the first, second, and M values after sorting the probabilities from high to low; C i，k is the result after standardizing all C i,n,k values; PC i，k is the calculated activity factor score; k∈1…M, M is the total number of transcription factors in each cell in the deep network model, and n∈1…N, N is the number of cell peaks that can be read in each cell; C i,n,k is the operation of assigning values according to Z i,n,k and the top two values of the output coincidence probability.

[0043] The beneficial effects of the above technical solutions are: the transcription factor activity score obtained by the present scheme can more intuitively find transcription factors with higher regulation strength or more regulation relationships. BRIEF DESCRIPTION OF DRAWINGS

[0044] Figure 1 is a flowchart of a single-cell transcription factor prediction method based on deep learning and attention mechanism.

[0045] Figure 2 is a framework diagram of a single-cell transcription factor prediction method based on deep learning and attention mechanism. DETAILED DESCRIPTION

[0046] The specific embodiments of the present application are described below to enable those skilled in the art to understand the present application, but it should be clear that the present application is not limited to the scope of the specific embodiments, and for those skilled in the art, it is obvious that various changes are within the spirit and scope of the present application defined and determined by the appended claims, and all the inventions utilizing the concept of the present application are within the scope of protection.

[0047] Reference Figure 1 , Figure 1 A flow chart of a single-cell transcription factor prediction method based on deep learning and attention mechanism is shown, as shown in Figure 1 , the method comprises steps S1 to S3.

[0048] In step S1, the single-cell chromatin accessibility analysis sequencing data is obtained and preprocessed, and then data enhancement operation is performed to obtain enhanced sequencing data;

[0049] In implementation, the method for preprocessing single-cell chromatin accessibility analysis sequencing data preferably comprises:

[0050] The peak read information of the single-cell chromatin accessibility analysis sequencing data is screened using the ENCODE single-cell chromatin accessibility analysis sequencing data processing method.

[0051] The bam file of the peak read information is converted to a bigwig file using deepTools2, and then all bigwig files are spliced into a data matrix using the bigWigMerge tool.

[0052] The data enhancement operation includes: calculating the potential characteristics of single cells using the cisTopic method, and calculating the similarity scores between cells using the cosine similarity data index; selecting 100 adjacent cells most similar to them as the data for amplification of single-cell chromatin accessibility analysis sequencing data, and collecting the single cells and the selected adjacent cells as enhanced sequencing data.

[0053] In step S2, the regression peak in the enhanced sequencing data is extracted to generate a 4*1000 feature vector S, the forward and reverse enhanced sequencing data is spliced into a 2*1000 feature vector A, and the DNA sequence data taken from the whole genome is converted into a 2*1000 feature vector U.

[0054] In step S3, the feature vector S, the feature vector A and the feature vector U are spliced into an 8*1000 feature vector, and a deep network model is input to predict the probability of each transcription factor in the single cell, the deep network model comprising a convolution module and a channel attention model.

[0055] The convolution layer in the convolution module mainly extracts the features of the input DNA sequence and shape data, mines the correlation between different data, removes noise and unstable components, and transmits the relatively stable information of the processed mode to the attention mechanism module for TFBS prediction.

[0056] The deep network model only uses the channel attention part in the CBAM to capture the features of the biological sequence and the shape data. The channel attention attention module compresses the input feature map in space. The compression method not only extracts features using average pooling, but also introduces maximum pooling as a supplement. Global average pooling has feedback on each pixel on the feature map, while global maximum pooling only has gradient feedback on the feature map with the largest response.

[0057] The present scheme can reduce noise in the data by preprocessing the single-cell chromatin accessibility analysis sequencing data, and then inputting the deep learning model embedded with the channel attention module for recognition, thereby improving the extraction speed and accuracy of the model for the required data features.

[0058] In an embodiment of the present application, the input deep network model further includes:

[0059] S31, judging whether the cell line to which the single-cell chromatin accessibility analysis sequencing data belongs is known, if yes, selecting the deep network model corresponding to the cell line for prediction, otherwise, entering step S32;

[0060] S32, inputting the spliced data into the deep network model corresponding to the cell line GM12878, K562 and H1ESC respectively for probability prediction of the transcription factor;

[0061] S33, selecting the maximum value of the transcription factor prediction probability output by the three models as the transcription factor prediction probability of the single cell.

[0062] The present scheme constructs a deep network model for a common cell line, so that the single cell can be quickly determined to be located in a specific cell line and the probability of the transcription factor in the cell can be accurately obtained even when the collected sequencing data information is limited.

[0063] As shown in Figure 2 The convolution module includes a convolution layer, an activation layer, a pooling layer and a full connection layer connected in sequence; the channel attention model includes a maximum pooling layer / average pooling layer, a shared multi-layer perception module, a full connection layer and a flattening layer connected in sequence; and the model structure of the deep network model is:

[0064] F1=max_pooling(ReLU(conv1(S FF2 = max_pooling(ReLU(conv2(F1))

[0065] F3 = max_pooling(ReLU(conv3(F2)), F = ReLU(W1 F3)

[0066] Z i,n,k = sigmoid(W2 M c (F))

[0067] where S F is the concatenated vector; ReLU(.) is the activation function; max_pooling(.) is the max pooling layer function; conv1(.), conv2(.) and conv3(.) are the first, second and third convolution functions, respectively; F1, F2 and F3 are the feature vector maps of each convolution layer of the convolution module, respectively; W1 and W2 are the weight matrices in the two fully connected layers, respectively; and are the features calculated by the global average pooling and the global maximum pooling in the channel attention model, respectively; σ is the sigmoid activation function; W0 and W3 are the two layers of parameters in the multilayer perceptron model, respectively; M c (F) is the output of the channel attention model; Z i,n,k is the coincidence probability of the transcription factor k and the nth cell peak in each cell i; k ∈ 1…M, M is the total number of transcription factors in each cell in the deep network model, and n ∈ 1…N, N is the number of cell peaks that can be read in each cell.

[0068] In implementation, the training method of the deep network model preferably comprises the following steps:

[0069] Preprocessing of DNA sequence data: selecting DNA sequence data of the whole genome, and cutting it into 200bp segments, with a sliding cutting interval of 50bp between each segment; obtaining the reverse strand of each segment, and expanding each reverse strand to a 1000bp segment with 200bp as the center, and then converting it into Mapability Data;

[0070] Preprocessing of batch organization chromatin accessibility analysis sequencing data: obtain batch organization chromatin accessibility analysis sequencing data in cell lines GM12878, K562, H1ESC, and perform shearing operation thereon, then map the sheared mutual data to human genome hg19 using Bowtie2 tool, and perform cutting using samtools and Picard; convert the cut bam file to bigwig file using deepTools2, and then splice all bigwig files into a data matrix using bigWigMerge tool;

[0071] Encode the DNA sequence data before preprocessing into a 4*1000 feature vector S1 using one-hot; splice the preprocessed batch organization chromatin accessibility analysis sequencing data corresponding to each cell line in the forward and reverse directions into a 2*1000 feature vector A1, A2 and A3; convert the Mapability Data into a 2*1000 feature vector U1;

[0072] Splice the feature vector S1, the feature vector A1 and the feature vector U1, the feature vector S1, the feature vector A2 and the feature vector U1, and the feature vector S1, the feature vector A3 and the feature vector U1, respectively, and input the spliced three vectors into three deep neural networks for training (the specific training process is input data-forward propagation-error calculation-backward propagation-parameter update-training end), to obtain the deep network model corresponding to the cell lines GM12878, K562 and H1ESC.

[0073] The single-cell transcription factor prediction method of the present scheme further comprises calculating the activity factor score of the transcription factor according to the probability of the transcription factor:

[0074]

[0075]

[0076]

[0077] wherein, are the coincidence probabilities of the transcription factor k and the nth cell peak in the cell i output by the final network model, and the probabilities are sorted from high to low as the first, second, and M values; C i，k is the result after standardizing all C i,n,k values; PC i，k is the calculated activity factor score; k∈1…M, M is the total number of transcription factors in each cell in the deep network model, and n∈1…N, N is the number of cell peaks that can be read in each cell; C i,n,k is the activity factor score according to Z i,n,kAn operation of assigning values to the top two values with the highest coincidence probability with the output.

[0078] In summary, the present scheme can effectively solve the problem of low data feature extraction efficiency caused by the black box characteristics of the neural network through the combination of the convolution module and the channel attention mechanism.

Claims

1. A method for single-cell transcription factor prediction based on deep learning and attention mechanism, characterized in that, The method comprises the steps of: obtaining single-cell chromatin accessibility analysis sequencing data, and preprocessing the data, and then performing data enhancement operation to obtain enhanced sequencing data; It comprises: The ENCODE single-cell chromatin accessibility analysis sequencing data processing method is used to screen peak read information of single-cell chromatin accessibility analysis sequencing data; The bam file of the peak read information is converted into a bigwig file using deepTools2, and then all bigwig files are spliced into a data matrix using a bigWigMerge tool; The data enhancement operation comprises: calculating the potential characteristics of the single cell by using a cisTopic method, and calculating the similarity score between cells by using a cosine similarity data index; selecting 100 adjacent cells most similar to the single cell as the data for amplification of the single-cell chromatin accessibility analysis sequencing data, and collecting the single cell and the selected adjacent cells as the enhanced sequencing data; Extracting regression peaks in sequencing data as feature vectors S Concatenating forward and reverse sequencing data as feature vectors A Converting DNA sequence data taken from a whole genome into feature vectors U ; Concatenated feature vector S , feature vector A and feature vector U , and input a deep network model to predict the probability of each transcription factor in a single cell, the deep network model comprising a convolution module and a channel attention model; S31, judge whether the cell line to which the single-cell chromatin accessibility analysis sequencing data belongs is known, if yes, select the deep network model of the corresponding cell line to perform prediction, otherwise, go to step S32; S32, input the spliced data into the deep network model corresponding to the cell lines GM12878, K562 and H1ESC respectively to perform probability prediction of transcription factors; S33, select the maximum value in the transcription factor prediction probability output by the three models as the transcription factor prediction probability of the single cell; The convolution module comprises a convolution layer, an activation layer, a pooling layer and a full connection layer connected in sequence; the channel attention model comprises a maximum pooling layer / average pooling layer, a shared multi-layer perception module, a full connection layer and a flattening layer connected in sequence; the model structure of the deep network model is: F 1= max _ pooling ( ReLU ( conv 1( S F )) ， F 2= max _ pooling ( ReLU ( conv 2( F 1)) F 3= max _ pooling ( ReLU ( conv 3( F 2))， F = ReLU(W 1 ·F 3 ) M c ( F ) = σ ( W 3( W 0( )) + W 3( W 0( )))， Z i,n,k = sigmoid ( W 2· M c ( F )) wherein, S F is the concatenated vector; ReLU . is an activation function; max pooling . is a max-pooling layer function; conv 1 . conv 2 . and conv 3 . are the first, second, and third convolutional layers, respectively; F 1 F 2and F 3are the feature vector maps of each convolutional layer of the convolutional module; W 1and W 2are the weight matrices of the two fully connected layers, respectively; and are the features computed by the global average pooling and global max pooling in the channel attention model, respectively; σ is a sigmoid activation function; W 0and W 3are the parameters of the two layers in the multi-layer perceptron model, respectively; M c F is the output of the channel attention model; Z i,n,k is the transcription factor k coincidence probability with the i th cell peak in each cell; n k ∈ 1… M , M is the total number of transcription factors in each cell in the deep network model, n ∈ 1… N , N is the number of cell peaks that can be read in each cell;​​​​​​​​​​ The training method of the deep network model comprises: Preprocessing of DNA sequence data: selecting the DNA sequence data of the whole genome, and cutting it into 200bp fragments, with a sliding cutting interval of 50bp between each segment; obtaining the reverse strand of each segment, and expanding each reverse strand to a 1000bp segment with 200bp as the center, and then converting it into Mapability Data; Preprocessing of bulk tissue chromatin accessibility analysis sequencing data: obtain bulk tissue chromatin accessibility analysis sequencing data in cell lines GM12878, K562, H1ESC, and perform shearing operation on the data, then map the sheared data to human genome hg19 using the Bowtie2 tool, and use samtools, Picard operation for cutting; convert the cut bam file to bigwig file using deepTools2, and then splice all bigwig files into a data matrix using bigWigMerge tool; Encode DNA sequence data before pre-processing into 4*1000 feature vector using one-hot S 1; concatenate forward and reverse pre-processed bulk tissue chromatin accessibility analysis sequencing data for each cell line as feature vector A 1, A 2 and A 3; convert Mapability Data into 2*1000 feature vector U 1; Concatenate feature vectors respectively S 1, feature vector A 1 and feature vector U 1, feature vector S 1, feature vector A 2 and feature vector U 1 and feature vector S 1, feature vector A 3 and feature vector U 1, the three vectors are input into three deep neural networks respectively for training, and the deep network models corresponding to the cell lines GM12878, K562 and H1ESC are obtained. Calculate the activity factor score of the transcription factor according to the probability of the transcription factor: ， wherein, are the transcription factors outputted by the last network model respectively k with the cell i the coincidence probability of the first n cell peak in the cell M , the first, second, until the value of the probability ranked from high to low; is the result after normalization of all values; C is the calculated activity factor score; i,n,k is the operation of assigning values according to Z i,n,k and the top two values of the output coincidence probability.

Images