Method and device for predicting single cell level open state of cis-regulatory element based on deep learning, and medium

Abstract

The invention discloses a method and equipment for predicting a single cell level open state of a cis-regulatory element based on deep learning, and a medium, and belongs to the technical field of bioinformatics. According to the application, a model for predicting the open probability of the cis-regulatory element in a single cell is constructed by utilizing single cell chromatin open sequencing data of a plurality of vertebrates based on a deep learning method. The open probability of the regulatory element in a single cell can be predicted by utilizing the model constructed by the invention, and the model is further applied to identifying a genome global regulatory signal, predicting a cross-species cell map based on a species genome sequence, analyzing a cross-species regulatory element sequence pattern, predicting a mutation site of a non-coding disease and predicting saturated mutation. The method has very important application value.

Description

Technical Field

[0001] The present application relates to the field of bioinformatics technology, and in particular to methods, devices and media for predicting the open state of cis-regulatory elements at the single-cell level based on deep learning. Background Art

[0002] Cells are the basic structural and functional units of organisms. Different types of cells share the same genome sequence and are controlled by cell-type-specific transcriptional programs to produce specific cell phenotypes. However, how the fate of each cell is determined by the same genome and how to perform precise transcriptional regulation based on the genome sequence are questions that have not yet been clearly answered.

[0003] Computational methods based on deep learning provide unprecedented perspectives and opportunities for interpreting the transcriptional regulation patterns of genomic sequences. These methods can directly extract potential regulatory information from the original sequence, and then predict the binding sites of transcription factors, epigenetic modification patterns, and gene expression levels. Large-scale epigenetic sequencing data have promoted the emergence of many deep learning models, such as Expecto, Basset, Basenji, and Enformer, which can achieve accurate prediction of chromatin open states to a certain extent. However, the establishment of these models is usually based on population-level data.

[0004] Traditional sequencing technology at the population level can only detect the average of intercellular signals, which often obscures the specific information of individual cells. Single-cell chromatin accessibility sequencing (scATAC-seq) can reveal the degree of chromatin accessibility in specific regions of individual cells at the single-cell level, thereby reflecting important information on transcriptional regulation such as transcription factor binding. Therefore, establishing a deep learning model at the single-cell level can accurately model the gene regulation of each cell type, thereby more deeply revealing the key molecular mechanisms in the process of cell fate determination. In addition, building a single-cell deep learning model across species in vertebrates can help us understand the principles of cell type evolution, providing valuable resources and new methods for studying regulatory languages ​​in different biological systems. Summary of the invention

[0005] In order to solve at least one of the above technical problems, the technical solution adopted in this application is as follows.

[0006] The first aspect of the present application provides a method for predicting the open state of a cis-regulatory element at the single-cell level based on deep learning, comprising the following steps: S101, obtaining single-cell chromatin openness sequencing data of multiple vertebrate species, and obtaining a matrix data set including cis-regulatory elements and their open states in each single cell, wherein the open states include "open" and "not open"; S102, randomly dividing the data set into a training set, a validation set, and a test set, and using a deep learning algorithm to construct a model for predicting the open probability of a cis-regulatory element in a single cell, wherein the model includes a two-dimensional convolutional layer, multiple residual layers, a bottleneck layer, and a final fully connected layer; S103, for the cis-regulatory element to be predicted, the DNA sequence of the cis-regulatory element is converted into a feature vector using the two-dimensional convolutional layer, and then the multiple residual layers are used to perform operations, and the output feature map is transformed by a linear activation function and finally output to the fully connected layer to predict the open probability of the cis-regulatory element in a single cell.

[0007] In some embodiments of the present application, the two-dimensional convolutional layer consists of 128 filters of size 7×1, followed by a batch normalization layer, a rectified linear unit, and a maximum pooling layer.

[0008] In some embodiments of the present application, the multiple residual layers include 4 layers, each layer has multiple blocks, and each block consists of two convolutional layers, two batch normalization layers and two rectified linear unit layers.

[0009] The second aspect of the present application provides a method for identifying a global regulatory signal of a genome, comprising the following steps: S11, based on the entire chromosome sequence of the whole genome, a window sliding is performed with a length of 500 bp and a step length of 100 bp. For the sequence of each sliding window, the open state in a single cell is predicted using any method described in the first aspect of the present application; S12, aggregate the single-cell level open probability of the whole genome to the cell type level to obtain the global regulatory signal of the genome.

[0010] The third aspect of the present application provides a method for predicting a cross-species cell atlas based on a species genome sequence, comprising the following steps: S21, sliding the whole genome sequences of the first vertebrate species and the second vertebrate species in a window with a length of 500 bp and a step length of 100 bp, and predicting the open state of the sequence in each sliding window in a single cell using any method described in the first aspect of the present application to obtain cis-regulatory element signals of the whole genomes of the first vertebrate species and the second vertebrate species; S22, based on the cis-regulatory element signals of the whole genome of the first vertebrate species and the second vertebrate species obtained in step S21, calling cis-regulatory elements at the whole genome level, thereby de novo predicting a set of potential cis-regulatory elements on the genomes of the first vertebrate species and the second vertebrate species; S23, based on the set of cis-regulatory elements obtained in step S22, predict the open states of these cis-regulatory elements in multiple lineages and multiple cell types.

[0011] A fourth aspect of the present application provides a method for analyzing cross-species regulatory element sequence patterns, comprising the following steps: S31, for multiple vertebrate species, using single nucleotide importance scores to generate specific sequence patterns, and obtaining important motifs using any method described in the first aspect of the present application; S32, quantifying the consistency of the important motifs obtained in step S31 with the motifs already in the database, and performing cluster analysis on the motifs of multiple vertebrate species.

[0012] The fifth aspect of the present application provides a method for predicting mutation sites of non-coding diseases, comprising the following steps: S41, obtaining a variation site, and taking the variation site as the center of the input sequence; S42, predicting the sequences before and after mutation respectively using any of the methods described in the first aspect of the present application to obtain the open probability, and determining the mutation effect of the variant site using the absolute value of the difference between the open probabilities obtained for the sequences before and after mutation.

[0013] A sixth aspect of the present application provides a method for predicting saturation mutations, comprising the following steps: S51, simulated saturation mutagenesis of the regulatory region of the gene, mutating the reference nucleotide to three other alternative nucleotide bases; S52, predicting the sequences before and after mutation respectively using any of the methods described in the first aspect of the present application to obtain the open probability, and determining the importance score of the original base where the saturation mutation is located using the absolute value of the difference between the open probabilities obtained for the sequences before and after the three mutations.

[0014] The present application provides a computer device in seven aspects, including: Memory for storing computer programs; A processor, configured to implement the steps of the method according to any one of claims 1 to 3 when executing the computer program.

[0015] In an eighth aspect, the present application provides a computer-readable storage medium, on which a computer program is stored. When the computer program is executed by a processor, the steps of any method described in the first aspect of the present application are implemented.

[0016] Compared with the prior art, this application has the following beneficial effects: This application uses the single-cell genome sequencing platform based on single-end adapter transposase independently developed by the laboratory to construct the whole-body single-cell chromatin openness map of five vertebrates: mice, chickens, geckos, salamanders, and zebrafish; and based on the cross-species chromatin openness map data set, a deep learning model is trained to provide a new deep learning method for predicting chromatin openness signals from a single-cell perspective. This method can be independent of epigenetic data, and can only predict the openness signal of cis-regulatory elements at the single-cell level based on the genome sequence; based on the training model, a prediction method for global regulatory signals comparable to epigenetic experiments is developed to predict global regulatory signals at the single-cell level; based on the single-cell chromatin openness data training model of multiple species, the similarity of the internal regulatory element pattern across vertebrate species is interpreted. The method provided in this application has very important guiding significance in terms of cell fate determination mechanism, disease risk site identification, and cross-species genome transcriptional regulation analysis.

[0017] It should be understood that the content described in this section is not intended to identify the key or important features of the embodiments of the present application, nor is it intended to limit the scope of the present application. Other features of the present application will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] By reading the detailed description below with reference to the accompanying drawings, the above and other purposes, features and advantages of the exemplary embodiments of the present application will become readily understood. In the accompanying drawings, several embodiments of the present application are shown in an exemplary and non-limiting manner, wherein: Figure 1 The whole-body single-cell chromatin accessibility maps of five vertebrates, namely mouse, chicken, gecko, salamander, and zebrafish, are constructed based on the single-end adaptor transposase-based single-cell genome sequencing platform in Example 1 of the present application.

[0019] Figure 2 A flowchart of Example 2 of the present application for establishing a prediction model based on deep learning and single-cell chromatin openness data is shown.

[0020] Figure 3 A benchmark comparison of input length and number of model residual network layers in the model building phase of Example 2 of the present application is shown.

[0021] Figure 4The figure shows the prediction accuracy of the model established in Example 2 of the present application in five vertebrates: mice, chickens, geckos, salamanders, and zebrafish.

[0022] Figure 5 The cell clustering consistency between the single-cell chromatin openness map predicted by the model established in Example 2 of the present application and the single-cell chromatin openness map obtained by actual observation is shown.

[0023] Figure 6 It shows that the whole genome level signal is obtained by using the model prediction in Example 3 of the present application.

[0024] Figure 7 The single-cell chromatin openness map and lineage distribution predicted by the model based on the human and macaque genomes in Example 4 of the present application are shown.

[0025] Figure 8 The correspondence between the motifs extracted using the model in Example 5 of the present application and the JASPAR database is shown.

[0026] Fig. 9 The tree clustering of erythroid-specific motifs in the cross-species model constructed in Example 5 of the present application is shown.

[0027] Fig.10 The disease-related fine-mapping sites with significant mutation effects predicted by the model in Example 6 of the present application are shown.

[0028] Fig.11 The disease-related fine-mapping sites with strong lineage-specific mutation effects predicted by the model in Example 6 of the present application are shown.

[0029] Fig.12 The prediction of the mutation effect of causal non-coding variation sites using the model in Example 6 of the present application is shown.

[0030] Fig.13 The key sequence pattern of the erythroid lineage for computer simulation of saturation mutation identification using the model in Example 7 of the present application is shown. DETAILED DESCRIPTION

[0031] Unless otherwise indicated, implied from the context, or customary in the prior art, all parts and percentages in this application are based on weight, and the tests and characterization methods used are all synchronized with the filing date of this application. Where applicable, the contents of any patent, patent application or disclosure involved in this application are fully incorporated herein by reference, and their equivalent patent families are also introduced as references, especially the definitions of relevant terms in the art disclosed in these documents. If the definition of a specific term disclosed in the prior art is inconsistent with any definition provided in this application, the definition of the term provided in this application shall prevail.

[0032] Numerical ranges in this application are approximate values, so unless otherwise specified, they may include values ​​outside the range. Numerical ranges include all values ​​from the lower limit to the upper limit in increments of 1 unit, provided that there is an interval of at least 2 units between any lower value and any higher value. For a range containing a value less than 1 or containing a fraction greater than 1 (e.g., 1.1, 1.5, etc.), 1 unit is appropriately considered to be 0.0001, 0.001, 0.01 or 0.1. For a range containing a single digit less than 10 (e.g., 1 to 5), 1 unit is generally considered to be 0.1. These are only specific examples of what is intended to be expressed, and all possible combinations of the values ​​between the lowest and highest values ​​listed are considered to be clearly recorded in this application.

[0033] The terms "comprising", "including", "having" and their derivatives do not exclude the presence of any other components, steps or processes, and are irrelevant to whether these other components, steps or processes are disclosed in this application. To eliminate any doubt, unless explicitly stated, all compositions using the terms "comprising", "including", or "having" in this application may include any additional additives, adjuvants or compounds. On the contrary, except for those necessary for operating performance, the term "essentially consisting of..." excludes any other components, steps or processes from the scope of any description of the term below. The term "consisting of..." does not include any components, steps or processes that are not specifically described or listed. Unless explicitly stated, the term "or" refers to the listed individual members or any combination thereof.

[0034] In order to make the technical problems, technical solutions and beneficial effects solved by the present application more clearly understood, the present application is further described in detail below in conjunction with embodiments.

[0035] The following examples are used to demonstrate the preferred embodiments of the present application. It will be understood by those skilled in the art that the techniques disclosed in the following examples represent techniques discovered by the inventors that can be used to implement the present application, and therefore can be considered as preferred embodiments of the present application. However, it should be understood by those skilled in the art based on this specification that many modifications can be made to the specific embodiments disclosed herein, and still the same or similar results can be obtained without departing from the spirit or scope of the present application.

[0036] Unless defined otherwise, 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.

[0037] Those skilled in the art will recognize, or be able to ascertain using no more than routine experimentation, many technical equivalents to the specific embodiments of the invention described herein. Such equivalents are intended to be encompassed by the claims.

[0038] The experimental methods in the following examples are conventional methods unless otherwise specified. The instruments and equipment used in the following examples are conventional laboratory instruments and equipment unless otherwise specified; the experimental materials used in the following examples are purchased from conventional biochemical reagent stores unless otherwise specified.

[0039] Example 1: Creating a model dataset based on multi-species single-cell chromatin openness maps In this example, five vertebrates, namely, mice, chickens, geckos, salamanders, and zebrafish, were selected to obtain single-cell chromatin accessibility data.

[0040] The inventors used a single-cell genome sequencing platform based on a single-end adaptor transposase to perform single-cell chromatin open sequencing on the whole body tissues of each species, and performed cell screening on the sequencing data: the transcription start site enrichment score (TSS enrichment score) and unique fragment number (UQ) of the single cell were used as quality control indicators, specifically: TSS enrichment score>10; unique fragment number>10000.

[0041] Finally, we obtained the single-cell chromatin openness map data of each species at the whole-body level, such as Figure 1 shown.

[0042] For each species, we downsampled to obtain a high-quality single-cell dataset with a cell number of about 50,000. At the same time, we filtered the cis-regulatory elements, requiring that each cis-regulatory element must be open in more than 1% of the cells. The number of cells and cis-regulatory elements for each species are shown in Table 1.

[0043] Table 1 Statistics of single-cell chromatin openness datasets of five species

[0044] In the process of making the model dataset, the open pattern of the cis-regulatory element was binarized into labels: 1 represents "open" and 0 represents "not open". Then, the cis-regulatory element sequence was encoded using one-hot encoding: 1000 represents base "A", 0100 represents base "C", 0010 represents base "G", and 0001 represents base "G".

[0045] Finally, the corresponding labels and inputs are paired to generate a complete dataset for deep learning: a sparse matrix with rows representing the sequences of cis-regulatory elements and columns representing the open labels of each single cell. For each species, the cis-regulatory element dataset is divided into three groups, including training set, validation set and test set. The inventors strictly exclude sequence overlap between the training set and the validation set and the test set.

[0046] For each species, all cis-regulatory elements on chromosome 8 were selected as the test set, while genes on other chromosomes were used for training and validation sets in a ratio of 9:1, respectively.

[0047] Example 2 Prediction of cis-regulatory element open signals at the single-cell level based on deep machine learning Based on the single-cell chromatin open maps of five species obtained in Example 1, this example develops a method for predicting the open signals of cis-regulatory elements at the single-cell level based on deep machine learning.

[0048] First, the constructed model architecture consists of several key components: (1) 2D convolutional layer It consists of 128 filters of size 7×1, followed by batch normalization, rectified linear unit (RELU), and max pooling layers.

[0049] (2) ResNext residual network module Contains 4 layers, each with a specific number of blocks (3, 4, 6, 3). Each block consists of two convolutional layers, two batch normalization layers, and two RELU layers.

[0050] (3) Dense bottleneck layer 32 units were set to serve as the dimensionally reduced representation of each cis-regulatory element, followed by a batch normalization layer and a dropout layer with a dropout rate of 0.2.

[0051] (4) Final fully connected layer The architecture was completed with the same number of cells as used. The openness probability of each cis-regulatory element in each cell was predicted using sigmoid.

[0052] The above carefully designed architecture reflects a clever combination of convolutional layers, residual layers, and dense layers, which enables the model to effectively capture complex patterns in DNA sequences. The last layer of the model uses a sigmoid activation function, which helps to generate predictions about the open signals of cis-regulatory elements at the single-cell level. The model structure is shown in the figure Figure 2 shown.

[0053] The inventors selected the best performing model by optimizing the input sequence length, the number of residual network layers, etc. Figure 3 As shown in the figure, the final model input sequence length is determined to be 500bp, and the number of residual network layers used in the model is 34.

[0054] The inventors used binary cross entropy loss (BCE Loss) as the objective function during the training process to quantify the difference between the prediction and the actual observation in each training cycle. The inventors used a model learning rate of 1e-3 and implemented an early stopping mechanism: the training was terminated when the minimum training loss did not decrease for 10 consecutive epochs to reduce the risk of overfitting.

[0055] The inventors specified that the selected training batch size was 100 and the Adam optimizer was used to update the model parameters. All models were implemented using PyTorch and the Anaconda3 distribution of Python and trained on a computer cluster equipped with four NVIDIA GTX1080Ti GPUs. The computation time varied depending on the characteristics of the dataset and the specific GPU device used during training to ensure efficient and effective model learning.

[0056] For models trained on single-cell chromatin open datasets of different species, the area under the receiver operating characteristic curve (AUROC) was used to measure model performance. For each single-cell task, the inventors evaluated the prediction performance by calculating the AUROC score and calculated the median AUROC for each cell lineage in each species-specific model. The results are shown in Figure 2. Figure 4 Shown In order to evaluate the inter-cell relationships in the predicted single-cell chromatin openness map, the inventors calculated the adjusted mutual information index (AMI) of the predicted and observed chromatin openness at the cell subtype level to quantify the consistency between the predicted cell map and the actual cell map. The results are shown in Figure 2. Figure 5 shown.

[0057] Example 3 Method for identifying global genomic regulatory signals First, the whole genome is divided into sequence windows of 500 bp each, with a step length of 100 bp. Then, the model of Example 2 of the present invention is used to predict the open signal of each 500 bp sliding window on the genome, and the open signal of each single cell predicted whole genome is aggregated to the cell type level.

[0058] like Figure 6 As shown, the prediction of the whole genome using the model of Example 2 is correlated with the observation data of traditional epigenetic experiments, reflecting the rationality of the model in identifying open signals at the whole genome level.

[0059] Example 4 De novo prediction of cross-species cell maps based on genome sequences First, the whole genome DNA sequences of humans and macaques were divided into 500 bp lengths and a step size of 100 bp. The model trained based on mouse data in Example 2 of the present application was used to predict the chromatin open probability of the corresponding sequences in the form of a sliding window, thereby obtaining the whole genome regulatory signal distribution of humans and macaques.

[0060] Next, the Macs2 tool was used to call cis-regulatory elements at the whole genome level of humans and macaques, with a calling threshold of 0.1, and a set of potential cis-regulatory elements on the human and macaque genomes was predicted from scratch.

[0061] Based on the set of cis-regulatory elements obtained by calling, the model can predict the openness of these cis-regulatory elements in multiple lineages and multiple cell types. The specific method is to forward pass the input cis-regulatory element DNA sequence within the model to calculate the output of the bottleneck layer as the potential low-dimensional embedding of the cis-regulatory element. Then, the inventors multiplied the embedding of the cis-regulatory element with the weight matrix extracted from the final fully connected layer by dot product to generate the openness of these cis-regulatory elements at the single-cell level, and obtained a single-cell chromatin open sparse matrix with row names as cis-regulatory elements and column names as cells.

[0062] Finally, the cell matrix was subjected to a series of dimensionality reduction and clustering operations in the Signac process, including dimensionality reduction by singular value decomposition (SVD) of the TF-IDF normalized matrix and clustering at a resolution of 0.2. For better visualization, the inventors used t-SNE embedding to create representative predicted cell maps for humans and macaques. Thus, the chromatin openness landscape of human and macaque cells predicted from scratch by the model was constructed, including multiple cell lineages, such as Figure 7 shown.

[0063] Example 5 Model-based explanatory exploration of regulatory element sequence patterns across species The TF-MoDISco tool was used to perform interpretative extraction on the model trained based on single-cell chromatin accessibility data of mice, chickens, geckos, salamanders, and zebrafish in Example 2 of the present invention.

[0064] The computational efficiency of TF-MoDISco was improved by randomly sampling 1000 cells and calculating the average significance score for each task. Then, the position frequency matrix of the motifs was extracted from the model according to TF-MoDISco, and the TomTom algorithm (v.5.3.3) was used to query the JASPAR2024 database to quantify the motif similarity, and the E-value threshold was set to 0.01, as shown in Figure 8 The inventors used the motifStack R package to visualize the above five vertebrate trained models, showing the clustering of erythroid cell-specific motifs on the genomes of different species, as shown in Fig. 9 shown.

[0065] Example 6 Model predicts non-coding disease-related sites Using the 341 reported disease-related fine-mapped variants as the center of the input 500bp sequence, the sequences before mutation (Ref) and after mutation (Alt) were input into the model trained with mouse chromatin openness data to predict the openness changes of the corresponding single cells before and after mutation. The mutation effect was defined as the absolute value of the difference between the predicted results before and after mutation. It was found that when the threshold of mutation effect>0.1 was used, the model could verify 265 disease-related variants. Among the variants with the most significant mutation effects, such as Fig.10 Compared with random mutations in nearby regions, the inventors observed strong lineage-specific mutation effects, such as Fig.11 For example, it was found that the rs2068888 (chr10:93079885:G>A) mutation would lead to upregulation of CYP26A1 in human hepatocytes, thus increasing cholesterol levels; and the rs34038797 (chr12:630843:C>G) mutation would lead to downregulation of NINJ2 in human immune cells, thus causing a decrease in lymphocyte counts.

[0066] The inventors then looked at the causal noncoding mutation sites that had been discovered and showed that the model accurately predicted the mutation effects of these functional sites. For example, the model predicted that -175 T>C and -202 C>G would lead to activation of HBG in the erythroid lineage, -86 C>A and -90 C>T would lead to downregulation of HBB in the erythroid lineage, and -20 T>A and -6 G>C would lead to downregulation of F9 in hepatocytes. Fig.12 shown.

[0067] The above results indicate that the model has a strong ability to predict the functional relevance of non-coding disease variants.

[0068] Example 7 Computer simulation of saturation mutation With the help of the model, in silico saturation mutagenesis (ISM) was performed on the 500bp promoter upstream sequence of the gene HBG1 to obtain the significance scores of all single nucleotides on the target sequence and screen important regulatory elements that may significantly affect gene expression.

[0069] For each base position on the 500bp sequence upstream of the HBG1 promoter, the inventors performed three forward passes of the model, mutating the reference nucleotide to three other alternative nucleotide bases. For each mutation, the inventors compared the alternative accessibility prediction with the reference accessibility prediction to calculate the difference in openness at the single-cell level. The inventors then normalized the ISM scores of the four nucleotides (A / T / C / G) at each position so that their sum was zero. The normalized ISM score at the reference nucleotide was taken as the importance score of the position, which indicates its contribution to openness.

[0070] By directly visualizing the importance score of each base in the 500 bp sequence upstream of HBG1 on the genome and scanning the sequence using FIMO, the inventors discovered important motifs that have been documented to affect erythroid transcriptional regulation, such as CCAATbox and TAL1 motifs (e.g. Fig.13 as shown).

[0071] All documents mentioned in this application are cited as references in this application, just as each document is cited as reference individually. In addition, it should be understood that after reading the above teaching content of this application, those skilled in the art can make various changes or modifications to this application, and these equivalent forms also fall within the scope defined by the claims attached to this application.

Claims

1. A method for predicting the open state of cis-regulatory elements at the single-cell level based on deep learning, characterized in that: The following steps are involved: S101, obtaining single-cell chromatin openness sequencing data of multiple vertebrate species, and obtaining a matrix data set including cis-regulatory elements and their open states in each single cell, wherein the open states include "open" and "not open"; S102, randomly dividing the data set into a training set, a validation set, and a test set, and using a deep learning algorithm to construct a model for predicting the open probability of a cis-regulatory element in a single cell, wherein the model includes a two-dimensional convolutional layer, multiple residual layers, a bottleneck layer, and a final fully connected layer; S103, for the cis-regulatory element to be predicted, the DNA sequence of the cis-regulatory element is converted into a feature vector using the two-dimensional convolutional layer, and then the multiple residual layers are used to perform operations, and the output feature map is transformed by a linear activation function and finally output to the fully connected layer to predict the open probability of the cis-regulatory element in a single cell.

2. The method according to claim 1, characterized in that The 2D convolutional layer consists of 128 filters of size 7×1, followed by a batch normalization layer, a rectified linear unit, and a max pooling layer.

3. The method according to claim 1, characterized in that The multiple residual layers include 4 layers, each layer has multiple blocks, and each block consists of two convolutional layers, two batch normalization layers and two rectified linear unit layers.

4. A method for identifying global regulatory signals of a genome, characterized in that: The following steps are involved: S11, based on the entire chromosome sequence of the whole genome, a window sliding is performed with a length of 500 bp and a step length of 100 bp, and for the sequence of each sliding window, its open state in a single cell is predicted using any method described in claims 1 to 3; S12, aggregate the single-cell level open probability of the whole genome to the cell type level to obtain the global regulatory signal of the genome.

5. A method for predicting cross-species cell maps based on species genome sequences, characterized in that: The following steps are involved: S21, sliding the whole genome sequences of the first vertebrate species and the second vertebrate species in a window with a length of 500 bp and a step length of 100 bp, and predicting the open state of the sequence in each sliding window in a single cell using the method described in any one of claims 1 to 3 to obtain cis-regulatory element signals of the whole genomes of the first vertebrate species and the second vertebrate species; S22, based on the cis-regulatory element signals of the whole genome of the first vertebrate species and the second vertebrate species obtained in step S21, calling cis-regulatory elements at the whole genome level, thereby de novo predicting a set of potential cis-regulatory elements on the genomes of the first vertebrate species and the second vertebrate species; S23, based on the set of cis-regulatory elements obtained in step S22, predict the open states of these cis-regulatory elements in multiple lineages and multiple cell types.

6. A method for analyzing cross-species regulatory element sequence patterns, characterized in that: The following steps are involved: S31, for multiple vertebrate species, using single nucleotide importance scores to generate specific sequence patterns, and obtaining important motifs using any method of claims 1 to 3; S32, quantifying the consistency of the important motifs obtained in step S31 with the motifs already in the database, and performing cluster analysis on the motifs of multiple vertebrate species.

7. A method for predicting mutation sites of non-coding diseases, characterized in that: The following steps are involved: S41, obtaining a variation site, and taking the variation site as the center of the input sequence; S42, predicting the sequences before and after mutation respectively using the method described in any one of claims 1 to 3 to obtain the open probability, and determining the mutation effect of the variant site using the absolute value of the difference in the open probability obtained from the sequences before and after mutation.

8. A method for predicting saturation mutation, characterized in that: The following steps are involved: S51, simulated saturation mutagenesis of the regulatory region of the gene, mutating the reference nucleotide to three other alternative nucleotide bases; S52, predicting the sequences before and after mutation respectively using the method described in any one of claims 1 to 3 to obtain the open probability, and determining the importance score of the original base where the saturation mutation is located using the absolute value of the difference in the open probability obtained from the sequences before and after the three mutations.

9. A computer device, characterized in that: include: Memory for storing computer programs; A processor, configured to implement the steps of the method according to any one of claims 1 to 3 when executing the computer program.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed by a processor, the steps of the method according to any one of claims 1 to 3 are implemented.

Images