A chromatin interaction detection method and system based on attention mechanism and deep learning

Abstract

The present application relates to a kind of chromatin interaction detection method and system based on attention mechanism and deep learning, comprising: constructing chromatin interaction detection model;Chromatin interaction detection model is trained;The DNA sequence to be detected is input into trained chromatin interaction detection model, chromatin interaction detection is carried out, and detection result is obtained.The present application can use DNA sequence data and functional genomic data to realize the prediction of chromatin interaction by calculation method without time-consuming and laborious and high cost biological experiment.In addition, the prediction method of the present application has very strong generalization ability in predicting the chromatin interaction of different cell lines, and only one model can be constructed to realize the prediction of the chromatin interaction of multiple cell lines without constructing multiple prediction models for the chromatin interaction of different cell lines, which greatly saves time cost and improves practicability.

Description

TECHNICAL FIELD

[0001] The application relates to a chromatin interaction detection method and system based on an attention mechanism and deep learning, and belongs to the technical field of bioinformatics. BACKGROUND

[0002] Chromatin interaction plays a key role in gene regulation, DNA replication, evolution and disease mechanism, and is of great help to understanding three-dimensional genome structure and analyzing three-dimensional chromatin patterns. Therefore, chromatin interaction detection has become a hot topic. In early research on chromatin interaction detection, biological methods such as orthogonal experiments such as Hi-C technology are usually used for detection. However, these technologies face problems such as high cost, time-consuming and labor-intensive. Subsequently, some computational methods for predicting chromatin interaction using DNA sequences were proposed to solve these problems. However, these computational methods have overfitting problems due to high overlap of data, and therefore the credibility of the prediction results is low, which brings great inconvenience to practical application.

[0003] In current research, some methods using machine learning predict chromatin interaction by combining sequence information and functional genomics signals. Deep learning methods learn the internal laws and representation levels of sample data, and are a new research direction in the field of machine learning. With the development of hardware computing power and big data, deep learning has been widely applied in the field of bioinformatics. The commonly used method is convolutional neural network (CNN). CNN extracts different features of the input through convolution calculation, and extracts more complex features through stacked convolution layers. In addition, the technology of natural language processing has been greatly promoted in the field of bioinformatics, and many studies use word vector technology to treat DNA sequences as sentences and k-mer as words to realize feature extraction. Therefore, the technology effectively improves the performance and generalization ability of predicting chromatin interaction by capturing data features through these deep learning neural networks and word vector technology. However, the performance and generalization of these methods are poor, and it is difficult to meet our requirements for high precision and high generalization of prediction. Therefore, it is an important research direction to realize high-precision and high-generalization prediction of chromatin interaction.

[0004] Chromatin interaction is crucial for gene transcription, regulation and expression, so detecting chromatin interaction is very important for the study of three-dimensional genome structure. Various computational methods for predicting chromatin interaction have been proposed in previous studies. However, existing methods do not consider the problem of data imbalance, resulting in poor performance and generalization ability of predicting chromatin interaction, which is difficult to meet the needs of various scenarios in reality. SUMMARY

[0005] In view of the deficiencies of the prior art, the present application provides a chromatin interaction detection method and system based on an attention mechanism and deep learning;

[0006] In reality, the prediction of chromatin interaction has problems of data set imbalance, low prediction accuracy and poor generalization performance. The problem of data set imbalance leads to difficulty in model training, and some evaluation indicators may cover up the problems of the model; the problem of low prediction accuracy leads to low credibility of the prediction result, making it difficult to analyze further; the problem of poor generalization performance leads to large differences in the prediction ability of chromatin interaction of different cell lines, so different prediction models need to be constructed for the chromatin interaction of different cell lines.

[0007] The present application uses DNA sequence data and functional genomic data to construct a high-precision and high-generalization model for predicting chromatin interaction based on an attention mechanism deep learning method, and evaluates the model through various indicators in unbalanced and balanced data sets, thereby effectively solving the problems existing in the prior art.

[0008] The present application improves the feature extraction of sequence information by using word vector technology, and then uses a convolutional neural network (CNN), an attention mechanism and a fully connected network to construct an effective and robust prediction model to realize the prediction of chromatin interaction of various cell lines. Finally, the prediction method proposed in the present research is compared with the prediction methods in previous researches. The analysis and comparison results show that the prediction model proposed in the present technology has the best performance and stability in predicting chromatin interaction, and has very strong generalization ability in predicting chromatin interaction of different cell lines.

[0009] Terminology explanation:

[0010] The word2vec model is a group of related models used to generate word vectors. These models are shallow and two-layer neural networks used to train to restructure linguistic word texts. The network is represented by words, and needs to guess the input words in adjacent positions. Under the bag-of-words model assumption in word2vec, the order of words is not important. After training, the word2vec model can be used to map each word to a vector, which can be used to represent the relationship between words. The vector is the hidden layer of the neural network.

[0011] The technical scheme of the present application is:

[0012] A chromatin interaction detection method based on an attention mechanism and deep learning, comprising:

[0013] Constructing a chromatin interaction detection model;

[0014] Training the chromatin interaction detection model;

[0015] inputting the DNA sequence to be detected into the trained chromatin interaction detection model to perform chromatin interaction detection and obtaining a detection result.

[0016] According to the application, the chromatin interaction detection model comprises three input ports, a convolution layer, a max pooling layer, a Dropout layer, a batch normalization layer, an attention layer and a full connection layer.

[0017] The convolution layer extracts nonlinear features of data; the max pooling layer enhances the robustness of the chromatin interaction detection model and avoids overfitting; the Dropout layer avoids overfitting by temporarily and randomly discarding neurons; the batch normalization layer accelerates the training of the chromatin interaction detection model while avoiding gradient disappearance; the attention layer is used to extract the most important part of sequence information; and the full connection layer performs classification and outputs a result.

[0018] The two inputs input1 and input2 of the chromatin interaction detection model are sequence information of two regions, and the sequence information is converted into word vector features through pre-training; the word vector features of the two regions are respectively subjected to the convolution layer, the max pooling layer and the Dropout layer, and then are horizontally spliced together; the batch normalization layer is used to accelerate the training of the chromatin interaction detection model while avoiding gradient disappearance; the attention layer is used to extract more important information; the result is spliced with input3 including functional genomic data, and the full connection layer is used for classification, and a prediction result is obtained according to a prediction probability value, if the prediction probability value is greater than 0.5, it is predicted that there is interaction, otherwise, it is predicted that there is no interaction.

[0019] Further preferably, the functional genomic data comprises Genomics features, Conservation scores, CTCF motif and Distances.

[0020] According to the application, the chromatin interaction detection model is trained, comprising:

[0021] The chromatin pairs in the data set are divided into a training set and a test set according to chromosome numbers, and ten-fold cross-validation is performed; in each round of training, the training set is used to train the chromatin interaction detection model, and the test set is used to detect the effect of the chromatin interaction detection model; in the training process of the chromatin interaction detection model, the loss value of the focal loss loss function is calculated to update the gradient, so as to realize the training and parameter updating of the chromatin interaction detection model.

[0022] Further preferably, in the chromatin interaction detection model training process, the parameters of the convolution layer are: 32 kernels, kernel size 50; the parameters of the max pooling layer are: pooling size 25, step size 25; the parameters of the Dropout layer are 0.5, that is, 50% of the neurons are temporarily discarded at each time; the parameters of the focal loss loss function are: weight parameter 0.75, adjustment factor 3.

[0023] According to the application, preferably, the DNA sequence to be detected is input into the trained chromatin interaction detection model for chromatin interaction detection to obtain a detection result, which includes:

[0024] First, the DNA sequences of the two regions to be detected are processed using word vector technology to obtain two word vector features;

[0025] Then, the two word vector features are respectively subjected to convolution calculation, max pooling operation and dropout layer, and then spliced, and processed using a batch normalization layer;

[0026] Thirdly, the attention mechanism is used to extract more important information in the sequence information and give it a higher weight;

[0027] Finally, the output result is spliced with the input including functional genomic data, and then classified through a full connection layer to determine the prediction result according to the prediction probability value.

[0028] Further preferably, the DNA sequence to be detected is processed using word vector technology, including: using all DNA sequences in the data set to construct a corpus, regarding each DNA sequence as a sentence, and dividing each four consecutive bases into a word; based on the corpus and the divided words, training a word2vec model to extract the features of the DNA sequence, and converting the DNA sequences of the two regions into word vector features.

[0029] Further preferably, the classification is performed through a full connection layer, and the prediction result is determined according to the prediction probability value, including: mapping the output to the final prediction probability through the full connection layer using relu and sigmoid as the activation function respectively, if the prediction probability exceeds 0.5, it is considered that the prediction result is chromatin interaction, otherwise it is considered that the prediction result is non-chromatin interaction.

[0030] According to the application, preferably, the loss function of the chromatin interaction detection model is focal loss, and the loss function FL(p t ) is as shown in formula (1):

[0031] FL(p t )=-α t (1-p t ) γ logpt (1)

[0032] In formula (1), alpha t is a weight, used to solve the problem of unbalanced data, gamma is a focal parameter, used to control the weight of difficult-to-classify samples, and p t As shown in formula (2):

[0033]

[0034] In formula (2), p is a predicted value, and y is a true label.

[0035] A chromatin interaction detection system based on an attention mechanism and deep learning comprises:

[0036] A chromatin interaction detection model construction module is configured to construct a chromatin interaction detection model.

[0037] A chromatin interaction detection model training module is configured to train the chromatin interaction detection model.

[0038] A detection module is configured to input a DNA sequence to be detected into the trained chromatin interaction detection model, perform chromatin interaction detection, and obtain a detection result.

[0039] A computer device comprises a memory and a processor, the memory stores a computer program, and the processor implements the steps of the chromatin interaction detection method based on the attention mechanism and deep learning when executing the computer program.

[0040] A computer-readable storage medium stores a computer program, and the computer program implements the steps of the chromatin interaction detection method based on the attention mechanism and deep learning when executed by a processor.

[0041] The present application has the following beneficial effects:

[0042] 1. The present application proposes a method for predicting chromatin interaction based on an attention mechanism deep learning method, which can extract word vector features of DNA sequences and combine functional genomic features, and then use a constructed deep learning model to predict chromatin interaction.

[0043] 2、The prediction method provided by the present application has very strong generalization ability in predicting the chromatin interactions of different cell lines, so only one model can be constructed to realize the prediction of the chromatin interactions of multiple cell lines without the need to construct multiple prediction models for the chromatin interactions of different cell lines, greatly saving time cost and improving practicability.

[0044] 3、The present application can be applied to the detection of chromatin interactions in biomedicine, so as to further study the corresponding three-dimensional genome structure, save a lot of money and time cost, and better analyze, treat and prevent diseases. BRIEF DESCRIPTION OF DRAWINGS

[0045] Fig. 1(a) is a schematic diagram of performance comparison of four models on GM12878 cell lines;

[0046] Fig. 1(b) is a schematic diagram of performance comparison of four models on IMR90 cell lines;

[0047] Fig. 1(c) is a schematic diagram of performance comparison of four models on K562 cell lines;

[0048] Fig. 1(d) is a schematic diagram of performance comparison of four models on HeLa-S3 cell lines;

[0049] Fig. 2(a) is a schematic diagram of the results of cross-cell line verification of the ERT model;

[0050] Fig. 2(b) is a schematic diagram of the results of cross-cell line verification of the XGBoost model;

[0051] Fig. 2(c) is a schematic diagram of the results of cross-cell line verification of the SGDC model;

[0052] Fig. 2(d) is a schematic diagram of the results of cross-cell line verification of the IChrom-Deep model;

[0053] Figure 3 The structure of the chromatin interaction detection model constructed by the present application is shown in the schematic diagram. DETAILED DESCRIPTION

[0054] The present application is further limited by the description and examples in the specification, but is not limited thereto.

[0055] Example 1

[0056] A chromatin interaction detection method based on attention mechanism and deep learning, comprising:

[0057] Constructing a chromatin interaction detection model;

[0058] Training the chromatin interaction detection model;

[0059] The DNA sequence to be detected is input into the trained chromatin interaction detection model for chromatin interaction detection to obtain a detection result.

[0060] Embodiment 2

[0061] The chromatin interaction detection method based on an attention mechanism and deep learning according to Embodiment 1 is different in that:

[0062] As shown in Figure 3 The chromatin interaction detection model includes three input ports, a convolution layer, a max-pooling layer, a Dropout layer, a batch normalization layer, an attention layer, and a fully connected layer.

[0063] The convolution layer extracts nonlinear features of data; the max-pooling layer strengthens the robustness of the chromatin interaction detection model and avoids overfitting; the Dropout layer avoids overfitting by temporarily and randomly discarding neurons; the batch normalization layer accelerates the training of the chromatin interaction detection model while avoiding gradient disappearance; the attention layer is used to extract the most important part of sequence information; and the fully connected layer performs classification and outputs a result.

[0064] The two inputs input1 and input2 of the chromatin interaction detection model are sequence information of two regions, and the sequence information is converted into word vector features through pre-training; the word vector features of the two regions are respectively subjected to the convolution layer, the max-pooling layer, and the Dropout layer, and then are horizontally spliced together; the batch normalization layer is used to accelerate the training of the chromatin interaction detection model while avoiding gradient disappearance; the attention layer is used to extract more important information; the result is spliced with input3 including functional genomic data, and the fully connected layer is used for classification, and a prediction result is obtained according to a prediction probability value, if the prediction probability value is greater than 0.5, it is predicted that there is interaction, otherwise, it is predicted that there is no interaction. The functional genomic data includes Genomics features, Conservation scores, CTCF motif, and Distances.

[0065] The present application further improves the prediction accuracy and generalization performance from the model level by improving the structure of the chromatin interaction detection model, and more comprehensively meets the requirements of practical applications.

[0066] The chromatin interaction detection model is trained, including:

[0067] The data set is from a previously published study. The chromatin pairs in the data set are divided into training set and test set according to chromosome number, and ten-fold cross validation is performed; that is, in each round of ten-fold cross validation, the training set and the test set come from completely different chromosome numbers. In each round of training, the training set is used to train the chromatin interaction detection model, and the test set is used to detect the effect of the chromatin interaction detection model; during the training of the chromatin interaction detection model, the loss value of the focal loss loss function is calculated to update the gradient, so as to realize the training and parameter updating of the chromatin interaction detection model.

[0068] During the training of the chromatin interaction detection model, the parameters of the convolution layer are: 32 kernels, kernel size is 50; the parameters of the max pooling layer are: pooling size is 25, step is 25; the parameters of the Dropout layer are 0.5, that is, 50% of the neurons are temporarily discarded at each time, which effectively avoids overfitting; the parameters of the focal loss loss function are: weight parameter is 0.75, and adjustment factor is 3.

[0069] The DNA sequence to be detected is input into the trained chromatin interaction detection model, and the chromatin interaction detection is performed to obtain the detection result, including:

[0070] Firstly, the DNA sequences of the two regions to be detected are processed using the word vector technology to obtain two word vector features; including: the length of the DNA sequence is 5000bp, so the dimension is high, which will contain more noise, which will greatly affect the accuracy of the model in predicting chromatin interaction. The present application uses all DNA sequences in the data set to construct a corpus, and each DNA sequence is regarded as a sentence, and every four consecutive bases are divided into a word; based on the corpus and the divided words, a word2vec model is trained to extract the features of the DNA sequence, and the DNA sequences of the two regions are converted into word vector features.

[0071] Then, the two word vector features are respectively subjected to convolution calculation, max pooling operation and dropout layer, and then spliced; the robustness of the model is improved, and overfitting is avoided. And using batch normalization layer for processing; aims to speed up the training speed and avoid gradient disappearance. Thirdly, the attention mechanism is used to extract more important information in the sequence information, and give it higher weight; complete further filtering.

[0072] Finally, the output result is spliced with the input including functional genomic data, and then classified through a full connection layer to judge the prediction result according to the prediction probability value. Including: the output is mapped to the final prediction probability through the full connection layer by respectively using relu and sigmoid as the activation function, if the prediction probability exceeds 0.5, it is considered that the prediction result is chromatin interaction, otherwise it is considered that the prediction result is non-chromatin interaction. The application further improves the prediction accuracy and generalization performance from the model level by improving the structure of the model, and more comprehensively meets the requirements of practical application.

[0073] The loss function of the chromatin interaction detection model is focal loss, which can effectively balance the contribution of positive and negative samples to the loss, and can make the model learn the characteristics of the data comprehensively. The loss function FL(p t ) is as shown in formula (1):

[0074] FL(p t )=-α t (1-p t ) γ logp t (1)

[0075] In formula (1), alpha t is a weight, used to solve the problem of unbalanced data, gamma is a focal parameter, used to control the weight of difficult-to-classify samples, and p t is as shown in formula (2):

[0076]

[0077] In formula (2), p is a prediction value, and y is a true label.

[0078] As a comparison, the performance of the method of the present application and the existing method in predicting chromatin interaction in the same cell line is evaluated, as shown in FIG. 1(a), FIG. 1(b), FIG. 1(c), FIG. 1(d), FIG. 1(a) is a schematic diagram of the performance comparison of four models on the GM12878 cell line; FIG. 1(b) is a schematic diagram of the performance comparison of four models on the IMR90 cell line; FIG. 1(c) is a schematic diagram of the performance comparison of four models on the K562 cell line; FIG. 1(d) is a schematic diagram of the performance comparison of four models on the HeLa-S3 cell line; wherein, from left to right, ERT, XGBoost, SGDC and the chromatin interaction detection model IChrom-Deep proposed by the present application are in turn the previous research. The abscissa is in turn the balanced accuracy (BACC), the accuracy (ACC), the Matthew correlation coefficient (MCC) and the F1 score; as can be seen from FIG. 1(a), FIG. 1(b), FIG. 1(c), FIG. 1(d), the chromatin interaction detection model IChrom-Deep proposed by the present application has a more outstanding performance compared with the existing model. The chromatin interaction detection model proposed by the present application is significantly higher than the previous research method in performance and stability.

[0079] Subsequently, the performance (F1 score) of the detection method of the present application and the previous research method in cross-cell line prediction is further compared, as shown in FIG. 2(a), FIG. 2(b), FIG. 2(c), FIG. 2(d). FIG. 2(a) is a schematic diagram of the results of the ERT model cross-cell line verification; FIG. 2(b) is a schematic diagram of the results of the XGBoost model cross-cell line verification; FIG. 2(c) is a schematic diagram of the results of the SGDC model cross-cell line verification; FIG. 2(d) is a schematic diagram of the results of the IChrom-Deep model cross-cell line verification; wherein, the ordinate is the cell line of the training set data source, and the abscissa is the cell line of the test set data source; for example, the abscissa is GM12878, and the ordinate is K562, which means that the K562 cell line data is used to train the model, and the F1 score index of the GM12878 cell line is predicted. As can be seen from FIG. 2(a), FIG. 2(b), FIG. 2(c), FIG. 2(d), the detection method of the present application has obvious advantages compared with the previous method, and no matter which cell line data set is used for training, the model has a more excellent effect. As can be seen from FIG. 2(a), FIG. 2(b), FIG. 2(c), FIG. 2(d), the heat map color representing the detection method of the present application is obviously deeper, which shows that the detection method of the present application has obvious improvement in prediction accuracy and generalization performance, and is more in line with the needs of practical application.

[0080] Embodiment 3

[0081] A chromatin interaction detection system based on attention mechanism and deep learning, comprising:

[0082] The chromatin interaction detection model construction module is configured to construct a chromatin interaction detection model.

[0083] The chromatin interaction detection model training module is configured to train the chromatin interaction detection model.

[0084] The detection module is configured to input the DNA sequence to be detected into the trained chromatin interaction detection model, perform chromatin interaction detection, and obtain a detection result.

[0085] Embodiment 4

[0086] A computer device includes a memory and a processor, the memory stores a computer program, and the processor implements the steps of the chromatin interaction detection method based on the attention mechanism and deep learning in embodiments 1 or 2 when executing the computer program.

[0087] Embodiment 5

[0088] A computer readable storage medium stores a computer program, and the computer program implements the steps of the chromatin interaction detection method based on the attention mechanism and deep learning in embodiments 1 or 2 when executed by a processor.

Claims

1. A method for detecting chromatin interaction based on attention mechanism and deep learning, characterized in that, The method comprises the following steps: constructing a chromatin interaction detection model; training the chromatin interaction detection model; inputting a DNA sequence to be detected into the trained chromatin interaction detection model to detect chromatin interaction and obtain a detection result; the chromatin interaction detection model comprises three input ports, a convolution layer, a max pooling layer, a Dropout layer, a batch normalization layer, an attention layer and a full connection layer; the convolution layer extracts nonlinear features of data; the max pooling layer strengthens the robustness of the chromatin interaction detection model and avoids overfitting; the Dropout layer avoids overfitting by temporarily and randomly discarding neurons; the batch normalization layer accelerates the training of the chromatin interaction detection model and avoids gradient disappearance at the same time; the attention layer is used to extract the most important part of sequence information; the full connection layer classifies and outputs a result; two inputs input1 and input2 of the chromatin interaction detection model are sequence information of two regions, and the sequence information is converted into word vector features through pre-training; the word vector features of the two regions are respectively subjected to the convolution layer, the max pooling layer and the Dropout layer, and then are horizontally spliced together; the batch normalization layer is used to accelerate the training of the chromatin interaction detection model and avoid gradient disappearance at the same time; the attention layer is used to extract more important information; the result is spliced with input3 including functional genomic data, and the full connection layer is used for classification to obtain a prediction result according to a prediction probability value; if the prediction probability value is greater than 0.5, it is predicted that there is interaction, otherwise, it is predicted that there is no interaction; the DNA sequence to be detected is inputted into the trained chromatin interaction detection model to detect chromatin interaction and obtain a detection result, which comprises the following steps: firstly, word vector technology is used to process DNA sequences of two regions to be detected to obtain two word vector features; then, the two word vector features are spliced after being subjected to convolution calculation, max pooling operation and Dropout layer respectively; and the batch normalization layer is used for processing; thirdly, the attention mechanism is used to extract more important information in sequence information and give higher weight to the information; finally, the output result is spliced with input including functional genomic data, and then is subjected to full connection layer classification to determine a prediction result according to a prediction probability value.

2. The method of claim 1, wherein the method is based on attention mechanism and deep learning. training the chromatin interaction detection model comprises the following steps: chromatin pairs in a data set are divided into a training set and a test set according to chromosome numbers, and ten-fold cross validation is performed; in each round of training, the training set is used to train the chromatin interaction detection model, and the test set is used to detect the effect of the chromatin interaction detection model; in the training process of the chromatin interaction detection model, the loss value of a focalloss loss function is calculated to update gradients, so as to realize the training and parameter updating of the chromatin interaction detection model. 3.The method of claim 1, wherein, The DNA sequence to be detected is processed using a word vector technology, including: constructing a corpus using all DNA sequences in the data set, regarding each DNA sequence as a sentence, and dividing each four consecutive bases into a word; based on the corpus and the divided words, training a word2vec model to extract features of the DNA sequence, and converting the DNA sequences of the two regions into word vector features. 4.The method of claim 1, wherein, Classification is performed through a full connection layer, and the prediction result is determined according to the prediction probability value, including: mapping the output to the final prediction probability through the full connection layer using relu and sigmoid as the activation function respectively, and if the prediction probability is greater than 0.5, the prediction result is considered to be chromatin interaction, otherwise, the prediction result is considered to be non-chromatin interaction.

5. The method of claim 1-4, wherein, The loss function of the chromatin interaction detection model is focal loss, that is, FL(p t ) as shown in formula (1): FL(p t ) = -a t (1 - p t ) γ logp t (1) In formula (1), α t is a weight for solving the problem of unbalanced data, γ is a focal parameter for controlling the weight of difficult-to-classify samples, p t As shown in formula (2): In formula (2), p is a prediction value, and y is a true label. 6.A chromatin interaction detection system based on attention mechanism and deep learning, characterized in that, It includes: The chromatin interaction detection model construction module is configured to construct a chromatin interaction detection model; The chromatin interaction detection model training module is configured to train the chromatin interaction detection model; The detection module is configured to input the DNA sequence to be detected into the trained chromatin interaction detection model to perform chromatin interaction detection and obtain a detection result; The chromatin interaction detection model includes three input ports, a convolution layer, a maximum pooling layer, a Dropout layer, a batch normalization layer, an attention layer and a full connection layer; The convolution layer extracts the nonlinear features of the data; the maximum pooling layer enhances the robustness of the chromatin interaction detection model and avoids overfitting; the Dropout layer avoids overfitting by temporarily and randomly discarding neurons; the batch normalization layer accelerates the training of the chromatin interaction detection model while avoiding gradient disappearance; the attention layer is used to extract the most important part of the sequence information; The full connection layer classifies and outputs the result; The two inputs input1 and input2 of the chromatin interaction detection model are sequence information of two regions, and the sequence information is converted into word vector features through pre-training; the word vector features of the two regions are horizontally spliced after passing through the convolution layer, the maximum pooling layer and the Dropout layer; the batch normalization layer is used to accelerate the training of the chromatin interaction detection model while avoiding gradient disappearance; the attention layer is used to extract more important information; the result is spliced with the input input3 including functional genomic data, and the full connection layer is used for classification, and the prediction result is obtained according to the prediction probability value, if the prediction probability value is greater than 0.5, it is predicted to have interaction, otherwise, it is predicted to have no interaction; The DNA sequence to be detected is input into the trained chromatin interaction detection model to perform chromatin interaction detection and obtain a detection result, including: First, the DNA sequences of the two regions to be detected are processed using a word vector technology to obtain two word vector features; Then, the two word vector features are spliced after convolution calculation, maximum pooling operation and dropout layer; and processed using a batch normalization layer; Thirdly, the attention mechanism is used to extract more important information in the sequence information and give it a higher weight; Finally, the output result is spliced with the input including functional genomic data, and then classified through a full connection layer to determine the prediction result according to a prediction probability value. 7.A computer device, comprising a memory and a processor, wherein the memory stores a computer program, and the computer device is configured to perform the method according to any one of claims 1-6 when the computer program is executed by the processor. The processor implements the steps of the chromatin interaction detection method based on the attention mechanism and deep learning in any of claims 1-5 when executing the computer program.

8. A computer-readable storage medium having stored thereon a computer program, characterized in that The computer program is executed by the processor to implement the steps of the chromatin interaction detection method based on the attention mechanism and deep learning in any of claims 1-5.

Images