Cell-specific enhancer prediction method and system based on dna sequence data

Abstract

The application provides a cell-specific enhancer prediction method and system based on DNA sequence data, which comprises the following steps: sequence information is extracted by using two different feature encoding modes for DNA sequences, and the two different feature encodings correspond to obtained Dna2vec features and motif frequency features; when the Dna2vec feature encoding is used, a pre-trained DNA model provided in Dna2vec is used to index sequence encoding; when the motif frequency feature encoding is used, the number of TFBS motifs in each DNA sequence is extracted and is converted into a frequency; the Dna2vec features and the motif frequency features are fused to construct a deep learning model; and whether the DNA sequence data belongs to an enhancer is judged based on the deep learning model.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of information processing, and particularly relates to a cell-specific enhancer prediction method and system based on DNA sequence data. BACKGROUND

[0002] The statements in this section merely provide background information related to the present application and do not necessarily constitute prior art.

[0003] Enhancers are non-coding fragments in DNA sequences that can regulate gene expression and transcription. As a class of regulatory elements, enhancers control various cellular activities such as specific gene expression, cell growth and differentiation, and cell carcinogenesis. Mutation or abnormal expression of enhancers can disrupt the gene regulatory network, thereby affecting cell function, tissue development, and disease progression. Therefore, the identification of enhancers is crucial for studying gene expression and regulation.

[0004] In recent years, several computational methods for enhancer identification have been proposed, but the data sets on which these methods are based have two problems:

[0005] First, the enhancers in this data set are extracted as short sequences of fixed length (200bp), which makes it unknown whether these methods can be used for unequal length sequences and maintain superior performance on unequal length sequences.

[0006] In addition, this data set is a mixed general data set, while studies have shown that enhancers are cell-specific, and the enhancer prediction methods studied on general data sets are not strongly applicable.

[0007] Therefore, the existing method Enhancer-IF recognizes these deficiencies and extracts enhancers from eight cell lines, i.e., uses a benchmark data set to predict cell-specific enhancers, but this method still has the shortcomings of simple feature extraction scheme, poor performance in predicting cell-specific enhancers, and incomplete performance verification. SUMMARY

[0008] To overcome the deficiencies of the prior art described above, the present application provides a cell-specific enhancer prediction method based on DNA sequence data to identify cell-specific enhancers.

[0009] To achieve the above-mentioned purpose, one or more embodiments of the present application provide the following technical solutions:

[0010] In a first aspect, a cell-specific enhancer prediction method based on DNA sequence data is disclosed, comprising:

[0011] Sequence information is extracted using two different feature encoding methods for DNA sequences, and the two different feature encodings correspond to obtaining Dna2vec features and motif frequency features;

[0012] wherein, when the Dna2vec feature encoding is adopted, a pre-trained DNA model provided in the Dna2vec is used to index the sequence encoding;

[0013] When the motif frequency feature encoding is adopted, the number of TFBS motifs in each DNA sequence is extracted and converted into a frequency;

[0014] The Dna2vec feature and the motif frequency feature are fused to construct a deep learning model;

[0015] The deep learning model is used to determine whether the DNA sequence data belongs to an enhancer.

[0016] As a further technical solution, the Dna2vec is used to calculate the distributed representation of variable length k-mers in DNA sequences, which uses human genome sequences as a learning corpus, uses a continuous skip-gram model in word2vec for unsupervised training, and embeds k-mers into a continuous vector space.

[0017] As a further technical solution, when the continuous skip-gram model in word2vec is used for unsupervised training, the pre-training performance is optimal when 3 nucleotides are used as a word.

[0018] As a further technical solution, when the continuous skip-gram model in word2vec is used for unsupervised training, it is specifically:

[0019] Suppose the length of each DNA sequence is L, when k = 3, it is divided into (L-2) words, using the pre-trained model in Dna2vec, the first feature vector corresponding to each word is obtained, and the feature vectors of all words are connected to obtain the second feature vector.

[0020] As a further technical solution, an adaptive pooling operation is used to normalize the feature dimensions so as to input them into a deep learning model.

[0021] As a further technical solution, when the number of TFBS motifs in each DNA sequence is extracted, the position weight matrix of the motif is obtained from the database, which is used for sliding scale matching with the sequence data in the data set, and the specific matching rule is:

[0022] Suppose the length of the motif is L m , then the PWM is a matrix with 4 behaviors L m , column, which represents the corresponding score of each base when the base is A, C, G and T;

[0023] Suppose the length of the DNA sequence is L s, the sequence is divided into L m sub-sequence segments with a step of 1, thereby obtaining L s -L m +1 sub-sequence segments;

[0024] For each sub-sequence segment, the sum of the score of each base pair is taken as the final matching score and compared with the corresponding threshold score to determine whether the sub-sequence segment matches the motif.

[0025] As a further technical solution, the rule for comparing sub-sequence segments is as follows:

[0026] Wherein, the value of j is 0, 1, 2 and 3, corresponding to the base A, C, G and T in the sub-sequence segment; Q represents the matching score;

[0027] Assuming P represents the p-value threshold score (10 -4 ) of the corresponding motif, when Q>P, it is considered that the sub-sequence segment matches the motif;

[0028] By traversing each sequence, the number information of each TFBS motif in each DNA sequence is obtained.

[0029] As a further technical solution, the feature vector obtained by dividing the number of motifs by the length of each sequence is taken as the input of the deep learning model.

[0030] As a further technical solution, the deep learning model includes a Dna2vec module and a motif module.

[0031] The Dna2vec module and the motif module use two sequence encoding schemes, Dna2vec and motif frequency, as input respectively.

[0032] As a further technical solution, in the Dna2vec module, data alternately passes through three one-dimensional convolution layers and three maximum pooling layers to extract features from DNA sequences, wherein the convolution layer uses convolution calculation to extract complex features of the input, and the maximum pooling layer adopts a downsampling method to output the maximum value of each sub-region.

[0033] The extracted features pass through a dropout layer to temporarily delete some neurons from the network at random during the training phase to prevent overfitting.

[0034] As a further technical solution, in the motif module, data passes through three fully connected layers for dimension reduction.

[0035] The extracted features pass through a dropout layer.

[0036] The Dna2vec module generates a third feature vector after feature extraction, the motif module generates a fourth feature vector, the two feature vectors are connected into a one-dimensional vector, and a linear layer with one neuron and a "sigmoid" activation function is used to output a prediction value;

[0037] If the prediction value is greater than a set value, the sample is considered to be a positive sample, that is, the sequence is considered to be an enhancer; otherwise, the sample is considered to be a negative sample, that is, the sequence is considered to be a non-enhancer.

[0038] In a second aspect, a cell-specific enhancer prediction system based on DNA sequence data is disclosed, comprising:

[0039] The sequence information extraction module is configured to extract sequence information by using two different feature encoding methods for DNA sequences, and the two different feature encodings correspond to obtaining Dna2vec features and motif frequency features;

[0040] When the Dna2vec feature encoding is used, a pre-trained DNA model provided in Dna2vec is used to index the sequence encoding;

[0041] When the motif frequency feature encoding is used, the number of TFBS motifs in each DNA sequence is extracted and converted into a frequency;

[0042] The deep learning model construction module is configured to fuse the Dna2vec features and the motif frequency features to construct a deep learning model;

[0043] The enhancer judgment module is configured to judge whether the DNA sequence data belongs to an enhancer based on the deep learning model.

[0044] The above one or more technical solutions have the following beneficial effects:

[0045] The present application uses two sequence encoding schemes that can be used to extract sequence information: Dna2vec feature encoding and motif frequency feature encoding, and a deep learning model is constructed by fusing Dna2vec features and motif frequency features and used for predicting cell-specific enhancers. The prediction model proposed in the present application can accurately judge whether the sequence belongs to an enhancer through DNA sequence data, proving that the present application is superior to existing methods for predicting specific enhancers; the prediction model proposed in the present application performs superior performance on other multiple data sets (universal data sets, enhancer-promoter data sets), proving the robustness of the present application in practical applications.

[0046] The advantages of the additional aspects of the present application will be partially given in the following description, partially will become apparent from the following description, or will be understood by the practice of the present application. BRIEF DESCRIPTION OF DRAWINGS

[0047] The accompanying drawings, which form a part of this specification, are included to provide a further understanding of the application, and are incorporated in and constitute a part of this specification. The embodiments of the application, and their

[0048] Figure 1 A flowchart of the method of the embodiment of the application;

[0049] Figure 2 A structural schematic diagram of the model of the embodiment of the application;

[0050] Figure 3 A comparison diagram for verifying the performance of the combination module of the embodiment of the application;

[0051] Figure 4 A comparison diagram of the embodiment of the application and the cell-specific enhancer prediction model Enhancer-IF;

[0052] Figure 5 A diagram of the best performance test of the embodiment of the application Enhancer-MDLF on an independent test set;

[0053] Figure 6 A diagram of the embodiment of the application, in which enhancer sequences are used as positive samples, promoter sequences are used as negative samples, and 20% of the sequences in each cell line are used as an independent test set for testing. DETAILED DESCRIPTION

[0054] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0055] It should be noted that the terms used herein are only for the purpose of describing specific embodiments and are not intended to limit the exemplary embodiments according to the application.

[0056] In the case of no conflict, the embodiments in the application and the features in the embodiments can be combined with each other.

[0057] Embodiment one

[0058] The embodiment discloses a cell-specific enhancer prediction method based on DNA sequence data, as shown in the accompanying drawings Figure 1 The method comprises the following steps:

[0059] Two different feature encodings are used to extract sequence information for DNA sequences, and the two different feature encodings correspond to obtaining Dna2vec features and motif frequency features;

[0060] When the Dna2vec feature encoding is adopted, a pre-trained DNA model provided in the Dna2vec is used to index the sequence encoding;

[0061] When the motif frequency feature encoding is adopted, the number of TFBS motifs in each DNA sequence is extracted and converted into a frequency;

[0062] The Dna2vec feature and the motif frequency feature are fused to construct a deep learning model.

[0063] Based on the deep learning model, it is determined whether the DNA sequence data belongs to an enhancer.

[0064] Two key steps: sequence encoding and model construction.

[0065] The sequence encoding includes Dna2vec and motif frequency.

[0066] The Word2vec word embedding technology is widely used in bioinformatics, but when using word vector technology for sequence encoding, the corpus used to train the word vector usually only has one cell line data set, and the learned word vector only contains a small amount of information. In order to solve the above limitation, the present application uses a pre-trained DNA model provided in the previous study Dna2vec to index the sequence encoding, which is an existing model in the previous study, and is described as follows: The pre-trained DNA model is based on the word2vec word embedding model, using human genome sequences as the learning corpus, using the continuous skip-gram model in word2vec for unsupervised training, and embedding k-mers into a 100-dimensional continuous vector space.

[0067] Through experiments in the range of k=[3,8], the optimal value of k is determined to be 3, that is, the pre-training performance is optimal when 3 nucleotides are used as a word. Therefore, assuming that the length of each DNA sequence is L, when k=3, it is divided into (L-2) words, using the pre-trained model in dna2vec, a 100-dimensional feature vector corresponding to each word is obtained, and all word feature vectors are connected to obtain a feature vector with a dimension of 100×(L-2), which is used as the feature encoding of each DNA sequence, and then input into the Dna2vec module of the deep learning model to learn the information therein.

[0068] In addition, since the feature dimensions of the unequal length sequences in the data set are not equal, an adaptive pooling operation is used to normalize the feature dimension to 10000 dimensions, so as to input them into the deep learning model.

[0069] Motif frequency:

[0070] Transcription factors play an important role in gene transcription by directly binding to motifs in the genome. Previous studies have found some potential transcription factor binding sites (TFBS) in the DNA sequences bound in the enhancer region.

[0071] Therefore, the present embodiment proposes to extract the number of TFBS motifs in each DNA sequence, and to perform length normalization on it to eliminate the influence of sequence length on motif number. This is input into the motif module of the deep learning model as a coding scheme for DNA sequences. The specific process is as follows:

[0072] The position weight matrix (PWM) of the motif is obtained from the HOCOMOCO Human v11 database, which is used for sliding scale matching with the sequence data in this dataset to determine whether a certain motif is likely to exist in the DNA sequence. The matching result is used for feature coding of the sequence. The specific matching rules are as follows:

[0073] Assume the length of the motif is L m , then PWM is a matrix of behavior L m , column 4, which represents the corresponding score of each base when the base is A, C, G and T.

[0074] Assume the length of the DNA sequence is L s , divide the sequence into subsequence fragments of length L m , with a step size of 1, to obtain L s -L m +1 subsequence fragments.

[0075] For each subsequence fragment, the sum of the corresponding score of each base is taken as the final matching score and compared with the corresponding threshold score to determine whether the subsequence fragment matches the motif. If the subsequence fragment matches the motif, it is counted as 1. The number of each motif contained in each DNA sequence is obtained by traversing the sequence.

[0076] The rules for comparing subsequence fragments are as follows:

[0077]

[0078] Where j takes the values 0, 1, 2 and 3, corresponding to the bases A, C, G and T in the subsequence fragment; Q represents the matching score; L represents the length of the motif. Assume P represents the p-value threshold score (10 -4 ) of the corresponding motif, then when Q>P, it is considered that the subsequence fragment matches this motif.

[0079] By traversing each sequence, the number of each TFBS motif in each DNA sequence is obtained, which is a 401-dimensional feature Vcount However, it is unfair to use only the motif number information due to the different sequence lengths, so the feature vector V is obtained by dividing the motif number by the length of each sequence frequency As the input of the final deep learning model, the calculation is as follows:

[0080]

[0081] where L s represents the length of each sequence.

[0082] Model construction:

[0083] As shown in the accompanying Figure 2 Enhancer-MDLF is a deep learning model, including a Dna2vec module and a motif module, and the Dna2vec module and the motif module use the two sequence encoding schemes of Dna2vec and motif frequency as input, respectively.

[0084] In the Dna2vec module, useful sequence information is extracted from the feature vector encoded by Dna2vec, and the calculation cost is reduced by using max pooling, and the characteristics of enhancers are fully considered by multiple convolution and pooling. The data first alternately passes through three one-dimensional convolution layers (Conv1d) and three max pooling layers (Maxpool) to extract features from DNA sequences. Among them, the convolution layer uses convolution calculation to extract the complex features of the input, and the max pooling layer uses the downsampling method to output the maximum value of each sub-region, thereby improving the robustness of the model and preventing overfitting.

[0085] The three one-dimensional convolution layers are set with parameters kernel_num=64, kernel_size=7, stride=3, and the three max pooling layers are set with parameters pool_size=2.

[0086] Subsequently, the extracted features pass through the dropout layer with dropout=0.6, so as to temporarily delete some neurons from the network at random in the training stage to prevent overfitting.

[0087] In the motif module, feature information is extracted from the feature vector encoded by the motif, and the data first passes through three fully connected layers (dense) to reduce the dimension to 16 dimensions. Subsequently, the extracted features pass through the dropout layer with dropout=0.6.

[0088] In addition, ReLU is used as the activation function for both modules of the deep learning model in order to add a nonlinear component in the model and enhance the expression ability of the model.

[0089] After feature extraction, the Dna2vec module generates a 500-dimensional feature vector, while the motif module generates a 16-dimensional feature vector. The two feature vectors are concatenated into a 516-dimensional one-dimensional vector, and a linear layer with one neuron and a "sigmoid" activation function is used to output the predicted value. If the predicted value is greater than 0.5, the sample is considered to be a positive sample, that is, the sequence is considered to be an enhancer; otherwise, the sample is considered to be a negative sample, that is, the sequence is considered to be a non-enhancer.

[0090] The multi-module deep learning prediction model proposed in the present application can analyze and extract features from multiple angles through convolution, pooling and other operations, and then predict cell-specific enhancers according to the features. Compared with previous methods, the classification effect of the present application is very excellent and better than other methods.

[0091] The benchmark dataset used in the embodiment of the present application includes eight different cell lines: GM12878, HEK293, HMEC, HSMM, HUVEC, K562, NHEK and NHLF. The positive samples of the dataset are obtained by the following method: first, use EnhancerAtlas 2.0 to extract the enhancer positions of each cell type; then obtain their corresponding sequences by using the SeqTailor tool, and the sequence length of most samples is between 200-2000bp; finally, for each cell type, use the CD-HIT database with a sequence recognition threshold of 60% to exclude homologous sequences. The negative samples of the dataset are obtained by the following method: randomly select sequence fragments with a length greater than 2000bp to construct candidate sequences for negative samples; then, trim the candidate sequences of negative samples to ensure that their length distribution is similar to that of the positive samples of each cell type.

[0092] In order to verify whether the combined module can effectively improve the performance of the model, the present application uses six performance evaluation indicators (AUC, BACC, MCC, SN, SP, F1) to evaluate the performance of the Dna2vec module and the motif module, and compares their performance with that of the combined module based on the independent test set of each cell line. As shown in Table 1, the combined module has better performance than the Dna2vec module and the motif module in terms of AUC, BACC, MCC, SN, SP and F1. Figure 3 It can be seen that the Enhancer-MDLF can effectively improve the performance of the model by fusing two modules. Especially on the HMEC cell line, the combined module greatly improves the prediction performance compared with the two independent modules. In terms of BACC, the performance of the combined module is 12.06% and 14.25% higher than that of the Dna2vec module and the motif module, respectively.

[0093] To demonstrate the superiority of the proposed Enhancer-MDLF in predicting cell-specific enhancers, it is compared with the only existing cell-specific enhancer prediction model, Enhancer-IF. The performance of these two methods on the independent test set (Appendix Figure 4 A-D) and ten-fold cross-validation (Appendix Figure 4 E-H) is compared. Compared with Enhancer-IF, Enhancer-MDLF performs well on all five metrics for all cell lines, which indicates that the performance of Enhancer-MDLF is significantly better than that of Enhancer-IF, demonstrating the superiority and robustness of Enhancer-MDLF in predicting cell-specific enhancers.

[0094] To demonstrate the robustness of the present application, Enhancer-MDLF is applied to the universal dataset, which is widely used for enhancer identification tasks. The enhancer sequences in this dataset are extracted into short sequences with a length of 200 bp. To ensure comparability, the same training set, test set, and evaluation metrics as previous studies are used. As shown in Appendix Figure 5 , Enhancer-MDLF achieves the best performance on the independent test set, with MCC, SN, and ACC of 0.6067, 0.84, and 0.8025, respectively.

[0095] In addition, some studies have shown that enhancers and promoters have similar sequence structures. Therefore, we use the dataset from iPro-WAEL to explore whether Enhancer-MDLF can accurately distinguish enhancers and promoters. This dataset contains four different human cell lines, namely GM12878, HeLa-S3, HUVEC, and K562, where enhancer sequences are used as positive samples and promoter sequences are used as negative samples, and 20% of the sequences in each cell line are used as an independent test set. Appendix Figure 6 shows that Enhancer-MDLF can capture the unique information of the two regulatory elements with similar sequence structures, thereby accurately distinguishing promoters and enhancers. These results all demonstrate the universality and robustness of Enhancer-MDLF.

[0096] Example Two

[0097] The purpose of this embodiment is to provide a computer device comprising a memory, a processor, and a computer program stored on the memory and executable on the processor, wherein the processor implements the steps of the above method when executing the program.

[0098] Example Three

[0099] The purpose of this embodiment is to provide a computer-readable storage medium.

[0100] A computer readable storage medium having stored thereon a computer program which, when executed by a processor, performs the steps of the above method.

[0101] Embodiment Four

[0102] The purpose of this embodiment is to provide a cell-specific enhancer prediction system based on DNA sequence data, comprising:

[0103] The sequence information extraction module is configured to extract sequence information by using two different feature encodings for DNA sequences, the two different feature encodings corresponding to obtaining Dna2vec features and motif frequency features;

[0104] When using Dna2vec feature encoding, a pre-trained DNA model provided in Dna2vec is used to index sequence encoding;

[0105] When using motif frequency feature encoding, the number of TFBS motifs in each DNA sequence is extracted and converted into frequency;

[0106] The deep learning model construction module is configured to use Dna2vec features and motif frequency features to construct a deep learning model;

[0107] The enhancer judgment module is configured to judge whether the DNA sequence data belongs to an enhancer based on the deep learning model.

[0108] The steps and methods involved in the above embodiments two, three and four correspond to embodiment one, and the specific implementation can refer to the related description part of embodiment one. The term "computer readable storage medium" should be understood to include a single medium or multiple media of one or more instruction sets; it should also be understood to include any medium that can store, encode or carry instruction sets for execution by a processor and cause the processor to perform any method in the present application.

[0109] Those skilled in the art should understand that each module or step of the present application described above can be realized by a general computer device, alternatively, they can be realized by program codes executable by a computing device, so that they can be stored in a storage device for execution by a computing device, or they can be made into individual integrated circuit modules, or a plurality of modules or steps among them can be made into a single integrated circuit module to realize. The present application is not limited to any specific combination of hardware and software.

[0110] The above describes the specific embodiments of the present application in combination with the drawings, but is not a limitation on the protection scope of the present application. Those skilled in the art should understand that various modifications or variations made by those skilled in the art on the basis of the technical solutions of the present application without creative labor are still within the protection scope of the present application.

Claims

1. A cell-specific enhancer prediction method based on DNA sequence data, characterized by: include: Two different feature encoding methods were used to extract sequence information from DNA sequences, and the two different feature encoding methods yielded Dna2vec features and motif frequency features respectively. When using Dna2vec feature encoding, the pre-trained DNA model provided in Dna2vec is used to index sequence encoding. When using motif frequency feature encoding, the number of TFBS motifs in each DNA sequence is extracted and converted into frequency; A deep learning model is constructed by fusing Dna2vec features and motif frequency features; the deep learning model includes a Dna2vec module and a motif module. The Dna2vec module and the motif module use Dna2vec and motif frequency sequence encoding schemes as inputs, respectively. In the Dna2vec module, data is alternately passed through three one-dimensional convolutional layers and three max pooling layers to extract features from DNA sequences. In the sequence module, the data is dimensionality reduced through three fully connected layers; the extracted features are processed through a dropout layer; the Dna2vec module generates a third feature vector after feature extraction, the sequence module generates a fourth feature vector, the two feature vectors are concatenated into a one-dimensional vector, and the predicted value is output using a linear layer with one neuron and a "sigmoid" activation function. Determining whether DNA sequence data belongs to enhancers based on deep learning models.

2. The cell-specific enhancer prediction method based on DNA sequence data as described in claim 1, characterized in that, The Dna2vec is used to compute the distributed representation of variable-length k-mers in DNA sequences. It uses human genome sequences as a learning corpus and performs unsupervised training using a continuous skip-gram model in word2vec to embed k-mers into a continuous vector space. When using a continuous skip-gram model for unsupervised training in word2vec, pre-training with three nucleotides as a single word yields the best performance. When using a continuous skip-gram model for unsupervised training in word2vec, the specific steps are as follows: Let the length of each DNA sequence be L. When k=3, it is divided into L-2 words. Using the pre-trained model in Dna2vec, the first feature vector corresponding to each word is obtained. The feature vectors of all words are concatenated to obtain the second feature vector. Adaptive pooling is used to normalize the feature dimensions so that they can be fed into the deep learning model.

3. The cell-specific enhancer prediction method based on DNA sequence data as described in claim 1, characterized in that, When extracting the number of TFBS motifs in each DNA sequence, a motif position weight matrix is ​​obtained from the database and used for sliding scale matching with the DNA sequence data. Specific matching rules are as follows: Assume the length of the motif is L m Then PWM is a behavior L m A matrix of 4 columns represents the score of each base when the bases are A, C, G and T. Assume the length of the DNA sequence is L s Divide the sequence into segments of length L m The subsequence fragment with a step size of 1 is used to obtain L. s - L m +1 subsequence fragments; For each subsequence fragment, the sum of the scores corresponding to each base is used as the final matching score and compared with the corresponding threshold score to determine whether the subsequence fragment matches the motif.

4. The cell-specific enhancer prediction method based on DNA sequence data as described in claim 3, characterized in that, The rules for comparing subsequence segments are as follows: Where j takes values ​​of 0, 1, 2, and 3, corresponding to the bases A, C, G, and T in the subsequence fragment; Q represents the matching score; Assuming P represents the p-value threshold score of the corresponding motif, then when Q > P, the subsequence fragment is considered to match this motif; By traversing each sequence, the number of each TFBS motif in each DNA sequence was obtained; The feature vector obtained by dividing the number of motifs by the length of each sequence is used as the input to the deep learning model.

5. The cell-specific enhancer prediction method based on DNA sequence data as described in claim 1, characterized in that, In the Dna2vec module, convolutional layers use convolution calculations to extract complex features from the input, while max pooling layers use downsampling to output the maximum value of each sub-region. The extracted features are passed through a dropout layer so that some neurons are temporarily and randomly removed from the network during the training phase to prevent overfitting.

6. The cell-specific enhancer prediction method based on DNA sequence data as described in claim 1, characterized in that, In the motif module, if the predicted value is greater than the set value, the sample is considered a positive sample, that is, the sequence is considered an enhancer; otherwise, the sample is considered a negative sample, that is, the sequence is considered a non-enhancer.

7. A cell-specific enhancer prediction system based on DNA sequence data, characterized in that: include: The sequence information extraction module is configured to extract sequence information from DNA sequences using two different feature encoding methods, with the two different feature encodings corresponding to Dna2vec features and motif frequency features. When using Dna2vec feature encoding, the pre-trained DNA model provided in Dna2vec is used to index sequence encoding. When using motif frequency feature encoding, the number of TFBS motifs in each DNA sequence is extracted and converted into frequency; The deep learning model building module is configured to fuse Dna2vec features and motif frequency features to build a deep learning model; the deep learning model includes a Dna2vec module and a motif module. The Dna2vec module and the motif module use Dna2vec and motif frequency sequence encoding schemes as inputs, respectively. In the Dna2vec module, data is alternately passed through three one-dimensional convolutional layers and three max pooling layers to extract features from DNA sequences. In the sequence module, the data is dimensionality reduced through three fully connected layers; the extracted features are processed through a dropout layer; the Dna2vec module generates a third feature vector after feature extraction, the sequence module generates a fourth feature vector, the two feature vectors are concatenated into a one-dimensional vector, and the predicted value is output using a linear layer with one neuron and a "sigmoid" activation function. The enhancer determination module is configured to determine whether DNA sequence data belongs to enhancers based on a deep learning model.

8. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps of the method described in any one of claims 1-6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it performs the steps of the method described in any one of claims 1-6 above.

Images