Cell cycle prediction method and system based on single-cell hi-c data

Abstract

The application provides a cell cycle prediction method and system based on single-cell Hi-C data, relates to the field of cell cycle prediction, and specifically includes the following steps: preprocessing Hi-C data of a single cell to be predicted to obtain a chromosome contact matrix; based on interaction information of the single cell to be predicted and the chromosome contact matrix, three feature sets of different scales are extracted by extracting a single-cell chromosome three-dimensional structure feature, namely, a contact probability distribution and genome distance feature set, a Bin contact probability feature set and a small-domain contact probability feature set; the three extracted feature sets are input into a trained fusion prediction model, convolution and feature fusion are performed, and finally, a cell cycle prediction result is generated and output; the application newly proposes two feature sets and constructs a fusion prediction model to independently predict a cell cycle to which a single cell belongs, and solves the problems of inaccurate prediction and low performance.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of cell cycle prediction, and particularly relates to a cell cycle prediction method and system based on single-cell Hi-C 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] The cell cycle process involves a series of changes in substances in cells, including RNA, DNA and proteins; analyzing the relationship between the cell cycle and the state of various substances will help analyze cell cycle dynamics; therefore, how to determine the cell cycle is very important.

[0004] Determining the cell cycle phase based on single-cell Hi-C data helps analyze and study the changes in the spatial structure of chromosomes at different cell cycle phases, which is of great significance to reveal cell cycle dynamics; however, the existing prediction method mainly uses machine learning method to establish cell cycle pseudo-trajectory sequence, although the method of creating cell cycle pseudo-trajectory sequence can initially identify the same type of cell population according to the cycle order, but this method still has some shortcomings: (1) without external auxiliary means, it is difficult to directly determine the cycle of a single cell in the cycle result; (2) the boundary between adjacent cell cycles in the sequence is relatively blurred; therefore, the existing technology has the problems of inaccurate prediction and low performance. SUMMARY

[0005] In order to overcome the shortcomings of the prior art, the present application provides a cell cycle prediction method and system based on single-cell Hi-C data, two new feature sets are proposed, and a fusion prediction model is constructed to independently predict the cell cycle to which a single cell belongs, solving the problems of inaccurate prediction and low performance.

[0006] To achieve the above object, one or more embodiments of the present application provide the following technical solutions:

[0007] The first aspect of the present application provides a cell cycle prediction method based on single-cell Hi-C data;

[0008] The cell cycle prediction method based on single-cell Hi-C data comprises:

[0009] The Hi-C data of the single cell to be predicted is preprocessed to obtain a chromosome contact matrix;

[0010] Based on the interaction information and the chromosome contact matrix of the single cell to be predicted, the three-dimensional structure features of the single cell chromosome are extracted to obtain three feature sets of different scales: contact probability distribution and genomic distance feature set, Bin contact probability feature set and intra-domain contact probability feature set;

[0011] The three extracted feature sets are input into the trained fusion prediction model, convolution and feature fusion are performed, and the final cell cycle prediction result is generated and output.

[0012] Further, the preprocessing, in particular:

[0013] Based on the read pair site mapping file in the single cell Hi-C data, the interaction information based on fragment number in the single cell Hi-C data is converted into interaction information based on fragment position coordinates;

[0014] All chromosomes in the single cell are divided into multiple fragments according to a preset resolution, and each fragment is called a Bin;

[0015] According to the interaction information based on fragment coordinate identification, the interaction information is mapped into the corresponding Bin, and the corresponding chromosome contact matrix for each chromosome of the single cell is generated.

[0016] Further, the contact probability distribution and genomic distance feature set, in particular:

[0017] The interaction of two fragments is regarded as a contact, and the contact is distributed to multiple intervals divided according to the linear distance of the two fragments in the contact;

[0018] The probability of contact count in each interval is calculated to obtain the contact probability distribution and genomic distance feature set of the cell.

[0019] Further, the Bin contact probability feature set, in particular:

[0020] Based on the chromosome contact matrix, the contact probability of each Bin on the chromosome is calculated; by calculating the Bin contact probability of all chromosomes, the Bin contact probability feature set of the cell is generated.

[0021] Further, the small domain contact probability feature set, in particular:

[0022] The small domain is defined as a triangular region centered on the target Bin and bounded by first-order linear adjacent Bins;

[0023] Based on the chromosome contact matrix, the contact probability of each small domain is calculated; by calculating the small domain contact probability of all chromosomes, the small domain contact probability feature set of the cell is generated.

[0024] Further, the fusion prediction model includes a convolution module and a feature fusion module;

[0025] The convolution module generates a one-dimensional feature vector for each feature set through a one-dimensional convolution layer, a batch normalization layer, a max pooling layer and a Dropout layer.

[0026] The feature fusion module fuses the three one-dimensional feature vectors into a final cell feature, and performs stage scoring on the cell feature, and the highest score corresponds to the final cell cycle prediction result.

[0027] Further, the training of the fusion prediction model adopts a 5-fold cross-validation method, and each fold divides a preset proportion of the training set data into a validation set, and the loss on the validation set does not decrease in a continuous preset number of rounds, and the model training is terminated.

[0028] The second aspect of the present application provides a cell cycle prediction system based on single-cell Hi-C data.

[0029] The cell cycle prediction system based on single-cell Hi-C data comprises a data preprocessing module, a feature extraction module and a cycle prediction module.

[0030] The data preprocessing module is configured to preprocess the Hi-C data of the single cell to be predicted to obtain a chromosome contact matrix.

[0031] The feature extraction module is configured to extract single-cell chromosome three-dimensional structure features based on the interaction information and the chromosome contact matrix of the single cell to be predicted, and obtain three feature sets of different scales: a contact probability distribution and genome distance feature set, a Bin contact probability feature set and an intra-domain contact probability feature set.

[0032] The cycle prediction module is configured to input the three extracted feature sets into the trained fusion prediction model, perform convolution and feature fusion, and generate and output the final cell cycle prediction result.

[0033] The third aspect of the present application provides a computer readable storage medium having a program stored thereon, which, when executed by a processor, implements the steps of the cell cycle prediction method based on single-cell Hi-C data according to the first aspect of the present application.

[0034] The fourth aspect of the present application provides an electronic device comprising a memory, a processor and a program stored on the memory and executable on the processor, wherein the processor executes the program to implement the steps of the cell cycle prediction method based on single-cell Hi-C data according to the first aspect of the present application.

[0035] The above one or more technical solutions have the following beneficial effects:

[0036] (1) The application proposes two new methods for extracting feature sets from single-cell Hi-C data, namely Bin Contact Probability feature set (BCP feature set) and Small Intra-Contact Probability feature set (SICP feature set), and a deep learning prediction model is constructed by fusing the existing genomic distance connection probability distribution feature set (CDD feature set) and the two new feature sets proposed in the application to predict the cell cycle.

[0037] (2) The deep learning prediction model proposed in the application can independently determine the cell cycle to which a single cell belongs, determine the cycle to which a fuzzy cell between different cell cycles belongs, and thus determine the boundary between different cell cycles.

[0038] The advantages of the additional aspects of the application will be partially given in the following description, partially become obvious from the following description, or be known by the practice of the application. BRIEF DESCRIPTION OF DRAWINGS

[0039] The drawings accompanying the specification of the application form part of the specification and serve to further illustrate the application, the illustrative embodiments of the application and the description thereof serve to explain the application without constituting an inappropriate limitation thereof.

[0040] Figure 1 The method flowchart of the first embodiment.

[0041] Figure 2 The small domain construction schematic diagram of the first embodiment.

[0042] Figure 3 The model structure diagram of the first embodiment.

[0043] Figure 4 The performance comparison diagram of the three feature sets of the first embodiment.

[0044] Figure 5 The performance comparison diagram of the feature set fusion model of the first embodiment.

[0045] Figure 6 The method comparison result diagram of the first embodiment.

[0046] Figure 7 The system structure diagram of the second embodiment. DETAILED DESCRIPTION

[0047] The application will be further described below in combination with the drawings and embodiments.

[0048] Embodiment one

[0049] The embodiment discloses a cell cycle prediction method based on single-cell Hi-C data;

[0050] AsFigure 1 The cell cycle prediction method based on single-cell Hi-C data includes the following steps.

[0051] Step S1: Preprocess the Hi-C data of the single cell to be predicted to obtain a chromosome contact matrix.

[0052] Most single-cell Hi-C data calculation methods represent single-cell Hi-C data as a plurality of chromosome contact matrices according to a given resolution. However, generating a chromosome contact matrix requires specific information of interacting chromosome fragments in each cell, which is not directly included in the original single-cell Hi-C data. Therefore, the embodiment uses two files in the Hi-C data, a read pair site mapping file and a chromosome interaction pair file, to construct a chromosome contact matrix.

[0053] The read pair site mapping file contains three columns of data, "fend", "chr", and "coord". The "fend" column stores the fragment number of the contact point on the chromosome, which uniquely identifies the fragment. The "chr" is the number of the chromosome where the fragment is located. The "coord" is the position coordinate of the fragment on the chromosome.

[0054] The chromosome interaction pair file is contained in the chromosome interaction pair folder. Each cell corresponds to a chromosome interaction pair file, which stores the interactions between different fragments on the same chromosome and the interactions between different fragments on different chromosomes in the cell. Each chromosome interaction pair file contains three columns of data, "fend1", "fend2", and "count". The "fend1" and "fend2" columns store the sequence numbers of the two chromosome fragments that interact. The "count" column stores the strength of the interaction between the two fragments, i.e., the number of contact points, which means that the two interacting fragments must have contact, and the two fragments with contact must have interaction, and the strength of the interaction is represented by the number of contact points.

[0055] Therefore, based on the read pair site mapping file and the chromosome interaction pair file in the Hi-C data, the embodiment preprocesses the single cell A as follows:

[0056] (1) Based on the read pair site mapping file in the single-cell Hi-C data, convert the interaction information based on the fragment number in the chromosome interaction pair file to interaction information based on the fragment position coordinate, specifically as follows:

[0057] Based on the information on the read pair site mapping file, each chromosome interaction pair file is mapped from the format of ["fend1", "fend2", "count"] to the format of ["chr1", "coord1", "chr2", "coord2", "count"], wherein "chr1" represents the number of the chromosome where the fragment with the serial number "fend1" is located, "coord1" represents the position coordinate of the fragment with the serial number "fend1" on the chromosome chr1, "chr2" represents the number of the chromosome where the fragment with the serial number "fend2" is located, "coord2" represents the position coordinate of the fragment with the serial number "fend2" on the chromosome chr2, and "count" represents the number of contact points between the two fragments.

[0058] Finally, the interactions between different fragments on different chromosomes are removed, and only the interaction information between different fragments on the same chromosome is retained.

[0059] (2) The chromosomes in the single cell A are divided into a plurality of fragments according to a preset resolution R, and each fragment is called a Bin, and specifically:

[0060] Based on the preset resolution R, each chromosome is re-divided into a plurality of fragments, in order to distinguish from the original fragments in the Hi-C data, each newly divided fragment is called a Bin in the application, and then the number of the Bin where each fragment is located is calculated, and the calculation formula is as follows:

[0061]

[0062] Wherein, Bin i represents the number of the Bin where the fragment i is located, floor() represents the down rounding operation, coord i represents the position coordinate of the fragment i on the chromosome, and resolution represents the preset resolution, that is, the length of the newly divided fragment.

[0063] (3) According to the interaction information based on the fragment position coordinates, the interaction information is mapped into the corresponding Bin, and the corresponding chromosome contact matrix C is generated for each chromosome of the single cell A, and specifically:

[0064] After the Bin number of each fragment is calculated, the number of contact points between each pair of Bins is recalculated according to the interaction information based on fragment position coordinates obtained in step (1), and a Bin interaction information file corresponding to each chromosome is obtained, which contains three columns of data, i.e., "bin1", "bin2" and "count", wherein "bin1" and "bin2" store the serial numbers of the two Bins that have interaction, and the "count" column stores the number of contact points between the two Bins.

[0065] Finally, according to the Bin interaction information file corresponding to each chromosome, a chromosome contact matrix C n*n is generated for each chromosome, wherein n represents the number of Bins divided by the chromosome, and C i,j represents the number of contact points between Bin i and Bin j .

[0066] Step S2: Multi-scale feature set extraction

[0067] Based on the interaction information of the single cell to be predicted and the chromosome contact matrix, single cell chromosome three-dimensional structure features are extracted, and three feature sets of different scales are obtained, i.e., a contact probability distribution and genome distance feature set, a Bin contact probability feature set and a small domain contact probability feature set.

[0068] (1) Contact probability distribution and genome distance feature set, i.e., CDD feature set.

[0069] The CDD feature set is not extracted from the chromosome contact matrix C, but from the interaction information based on fragment position coordinates obtained in step (1) of step (1) of step S1, and the interaction information based on fragment position coordinates is in the format of ["chr1", "coord1", "chr2", "coord2", "count"].

[0070] According to the coordinate information in the interaction information, i.e., the position coordinates of the first fragment on the chromosome coord1 and the position coordinates of the second fragment on the chromosome coord2, the linear distance of the two fragments in each contact is calculated, i.e., coord1-coord2, and the contact of the two fragments is assigned to d intervals divided according to the linear distance, and the linear distance range of the d intervals gradually increases, and the formula for assignment is as follows:

[0071]

[0072] wherein interval locinterval number assigned, coord1 and coord2 represent the position coordinates of two interacting fragments on the chromosome, and s represents the index step of each interval, in this embodiment, the value of s is set to 0.125, and the value range of coord1-coord2 is [2K, 9.3M].

[0073] After all contacts are assigned, the contact probability distribution and the genomic distance feature set CDD are extracted by calculating the probability of contact count in each interval, and the specific formula is as follows:

[0074]

[0075] wherein, CDD(interval loc ) represents the contact probability of interval interval loc , Total(interval loc ) represents the contact count in interval interval loc , and Total(Cell) represents the total number of contacts in all chromosomes in the current cell.

[0076] The contact count in interval interval loc is the sum of the contact points of all contacts divided into interval interval loc , for example, a pair of interacting fragments has coord1 and coord2 of 10010 and 10110, and the contact point number is 3; another pair has coord1 and coord2 of 20010 and 20110, and the contact point number is 2, then coord1-coord2 of the two pairs is equal to 100, then interval interval loc calculated according to formula (1) is the same, so the two pairs of fragments are divided into the same interval, then the contact count of the interval is 3+2=5, and the implementation of the contact count is to add the contact point numbers of the two fragments meeting the requirements in the for loop.

[0077] Finally, the CDD features corresponding to each cell are taken as row data, and the CDD features of all cells form a CDD matrix as shown in Figure 1 , wherein the number of cells in the Hi-C data is m, and the CDD matrix has m rows.

[0078] (2) Bin contact probability feature set, namely BCP feature set.

[0079] The contact frequency distribution of the same fragment on the same chromosome in different cell cycles is different, so the BCP feature set is extracted from the overall chromosome interaction information, the Bin contact probability set of the chromosome is used as the contact feature set of the chromosome, and the contact feature set of all chromosomes in the cell is used as one of the three feature sets extracted by the application.

[0080] Specifically, assuming that the chromosome contact matrix is C n*n , n is the number of Bins that can be divided by the current chromosome, since different chromosomes have different lengths, the number of Bins that can be divided at the same resolution is also different, and the Bin contact probability of the chromosome is calculated as follows:

[0081]

[0082] Total(C n*n ) is the contact total of the chromosome where the Bin i is located, that is, the sum of each element in the chromosome contact matrix C i corresponding to the chromosome where the Bin n*n is located; C ij represents the contact number of the i-th Bin and the j-th Bin; wherein Total(C n*n ) is calculated by calculating the Bin contact probability of all chromosomes, and the specific calculation formula is:

[0083]

[0084] Through the above two formulas, the contact probability of all Bins of each chromosome of each cell is calculated, and the BCP feature set is obtained.

[0085] The SICP feature set is calculated by calculating the contact probability in the small domain (3), that is, the SICP feature set.

[0086] The SICP feature set is extracted from the perspective of chromosome small domain interaction information, the small domain is defined as a triangular region with the target Bin on the diagonal of the chromosome contact matrix as the center and the first-order linear adjacent Bin as the boundary, and the small domain corresponding to the Bin k is shown in Figure 2 , in this embodiment, the triangular region composed of (k-1, k-1), (k-1, k), (k-1, k+1), (k, k), (k, k+1), (k+1, k+1) is the small domain with Bin k as the center.

[0087] Similar to the Bin contact probability, the contact probability in the small domain is calculated on the chromosome contact matrix C n*n , but before calculating the SICP feature, a circle of 0 is added outside the matrix, so that Cn*n becomes C (n+2)*(n+2) , so that k takes 1 and n can also constitute a complete small domain, the chromosome small domain contact probability is calculated as follows:

[0088]

[0089] Where, Bin k Indicates the center of the current calculation small domain, the value range of k is [1, n]. The small domain contact probability of all chromosomes

[0090] Through formula (5), the contact probability of all small domains of each chromosome of each cell is calculated, and the SICP feature set is obtained.

[0091] Need to be particularly pointed out is:

[0092] Different lengths of different chromosomes, according to the same resolution R, the number of Bin n is also different, so the data dimension of BCP feature set and SICP feature set is different.

[0093] Step S3: input the three feature sets extracted into the trained fusion prediction model, perform convolution and feature fusion, generate and output the final cell cycle prediction result.

[0094] In order to accurately and efficiently predict the cell cycle, the present application establishes a fusion prediction model based on deep learning, each row in the CDD feature set generated in step S2 is the CDD feature data of a cell, and the BCP and SICP feature sets need to combine the feature data corresponding to all Bin on all chromosomes in each cell into a one-dimensional vector, and each one-dimensional vector corresponds to the feature data of a cell. Figure 3 is a structural diagram of the fusion prediction model, which is drawn taking the data corresponding to a cell as an example, as shown in Figure 3 The fusion prediction model includes a convolution module and a feature fusion module.

[0095] The same network model is created for the three feature sets of CDD, BCP and SICP in the convolution module, specifically including one-dimensional convolution layer (Conv1d), batch normalization layer (BatchNorm), maximum pooling layer (MaxPool) and Dropout layer connected in turn; more complex features are obtained by stacking the above layers twice. The one-dimensional convolution layer (Conv1d) uses a one-dimensional convolution kernel with kernel_size = 7 and channel = 32 to collect features from various input feature sets; the batch normalization layer (BatchNorm) is used to prevent gradient explosion and gradient disappearance; the maximum pooling layer (MaxPool) is used to reduce the feature dimension, retain key features, reduce model calculation, avoid overfitting and enhance generalization; in addition, ReLU is used as an activation function to connect the BatchNorm layer and the MaxPool layer, so as to add a nonlinear component in the model and enhance the expression ability of the model; the Dropout layer can discard part of the neurons during the forward propagation process with a pre-set probability, thereby effectively preventing the model from overfitting; finally, the data generated from the network model is subjected to a Flatten operation to combine the data from all channels into a one-dimensional feature vector.

[0096] In the feature fusion module, the three one-dimensional feature vectors corresponding to the three feature sets generated by the convolution module are combined into a final cell feature, and then the scores of each stage are mapped into the cell feature using a linear layer and a "log_softmax" function. The final cycle prediction result is the stage corresponding to the highest score; using the "log_softmax" function not only avoids the value overflow problem, but also facilitates the calculation of the cross-entropy loss of the model.

[0097] The Nagano dataset from the GEO database is used as the benchmark dataset for studying the technical solutions of the present application (Nagano et al. 2017), and the Nagano dataset is downloaded from GSE94489 in the GEO database; after batch sequencing Hi-C data of mouse embryonic stem cells, Nagano et al. labeled 1171 mouse cells in the cell cycle stage using fluorescence-activated cell sorting technology; the cell cycles of these mouse embryonic stem cells are as follows:

[0098] 280 cells are in the G1 stage;

[0099] 303 cells are in the early-S stage;

[0100] 262 cells are in the mid-S stage;

[0101] 326 cells are in the late-S / G2 stage.

[0102] The cells in the benchmark dataset undergo data preprocessing in step S1 and multi-scale feature set extraction in step S2 to obtain three feature sets for each cell. The fusion prediction model is trained using the three feature sets of each cell as input and the cell's stage as the label.

[0103] The model training process includes an "early stop" mechanism and a 5-fold cross-validation method, in which 20% of the training data is split into the validation set in each fold to prevent the model from overfitting to the training set. If the loss on the validation set does not decrease in 10 consecutive rounds, the model training is terminated.

[0104] The original Nagano dataset contains read-pair mapping files and chromosome interaction pair files. The cells in the original Nagano dataset are subjected to the data preprocessing operation in step S1, which generates a specific chromosome interaction information file for each chromosome in the cell and maps the interaction information in the chromosome interaction information file to the corresponding Bin.

[0105] This invention establishes a deep learning-based fusion prediction model for multi-scale feature sets. It analyzes and fuses features from multiple perspectives through operations such as convolution and pooling, and then predicts the cell cycle stage based on the features. Compared with previous methods, the cycle prediction effect of this invention is excellent and superior to other methods.

[0106] To demonstrate the effectiveness of the multi-scale contact probability feature set, this invention verifies and analyzes the performance of the extracted feature set: First, the convolutional module of each feature set is included in the fusion prediction model and the corresponding feature set is input; then, the prediction performance of the three feature sets is evaluated using three evaluation metrics (ACC, F1, Precision). Figure 4 In the graph, A is the overall performance comparison chart of the feature set. Figure 4 It can be observed that each feature set is effective for predicting the four cell cycles.

[0107] This invention evaluates the performance of the three feature sets used in the model for predicting different cell cycles; Figure 4 In the graph, B is a comparison of different cell cycle performances of the feature set, and the x-axis represents the cell cycle. Figure 4In B, we can find that: (1) in G1 phase, the results of the three feature sets are similar; (2) in early-S phase, the results of BCP feature set are slightly better than CDD feature set, and the results of SICP feature set are the lowest; (3) in mid-S phase, the results of SICP and BCP feature sets are significantly greater than CDD feature set; (4) in late-S phase, the results of the three feature sets are close; these results show that the dynamic changes of chromosome three-dimensional structure are more dramatic in G1 phase and early-S phase, while the changes of chromosome structure are gradual and smooth from mid-S phase to late-S phase; in addition, from the difference between CDD feature set and the other two feature sets, it can be inferred that the structural changes of different chromosomes are more obvious in mid-S and late-S phases.

[0108] To evaluate the effectiveness of the fusion prediction model based on multiple feature sets of the present application, the present application compares the performance of the model with the fusion of the three feature sets with other six models (three single feature set models: CDD, BCP and SICP; three double feature set fusion models: CDD-BCP, CDD-SICP and BCP-SICP). Based on the three evaluation indexes of ACC, F1 and Precision, the comparison results are shown in Figure 5 It is obvious that the fusion prediction model proposed by the present application successfully fuses feature sets of different scales and achieves excellent cell cycle prediction performance.

[0109] Finally, to prove the superiority of the cell cycle prediction method based on single-cell Hi-C data proposed by the present application, the present application compares it with two cell cycle trajectory construction methods (Nagano and Circlet) using four evaluation indexes; in addition, the present application uses three traditional machine learning methods (support vector machine, logistic regression and random forest) to further evaluate the effectiveness of the deep learning model constructed by the present application; in terms of ACC, F1, Precision and ARI indexes, the comparison results are shown in Figure 6 It can be seen from Figure 6 that the performance of the multi-feature set fusion prediction model based on deep learning method is better; although Circlet has good performance in predicting cell cycle, it is not as good as the method of the present application in terms of ACC (0.854 vs 0.8766) F1 (0.8517 vs 0.8756), Precision (0.8517 vs 0.8757), ARI (0.6596 vs 0.6998) and the like; the comparison of these results shows that the cell cycle prediction method based on single-cell Hi-C data provided by the present application achieves the best performance in cell cycle prediction.

[0110] Example Two

[0111] The embodiment discloses a cell cycle prediction system based on single-cell Hi-C data;

[0112] As shown in the figure, the cell cycle prediction system based on single-cell Hi-C data comprises a data preprocessing module, a feature extraction module and a cycle prediction module: Figure 7

[0113] The data preprocessing module is configured to preprocess the Hi-C data of the single cell to be predicted to obtain a chromosome contact matrix.

[0114] The feature extraction module is configured to extract single-cell chromosome three-dimensional structure features based on the interaction information and the chromosome contact matrix of the single cell to be predicted, to obtain three feature sets of different scales: a contact probability distribution and genome distance feature set, a Bin contact probability feature set and an intra-domain contact probability feature set.

[0115] The cycle prediction module is configured to input the three extracted feature sets into a trained fusion prediction model, perform convolution and feature fusion, and generate and output the final cell cycle prediction result.

[0116] Embodiment three

[0117] The purpose of the embodiment is to provide a computer-readable storage medium.

[0118] The computer-readable storage medium stores a computer program, which is executed by a processor to implement the steps in the cell cycle prediction method based on single-cell Hi-C data according to the embodiment one of the present disclosure.

[0119] Embodiment four

[0120] The purpose of the embodiment is to provide an electronic device.

[0121] The electronic device comprises a memory, a processor and a program stored in the memory and executable on the processor, and the processor executes the program to implement the steps in the cell cycle prediction method based on single-cell Hi-C data according to the embodiment one of the present disclosure.

[0122] The above only describes the preferred embodiments of the present application and is not intended to limit the present application. Those skilled in the art can make various modifications and changes to the present application. Any modification, equivalent replacement, improvement, etc. made within the spirit and principles of the present application shall be included in the protection scope of the present application.​

Claims

1. A cell cycle prediction method based on single-cell Hi-C data, characterized in that, The method comprises the following steps: Preprocessing Hi-C data of a single cell to be predicted to obtain a chromosome contact matrix; Based on the interaction information of the single cell to be predicted and the chromosome contact matrix, single cell chromosome three-dimensional structure features are extracted to obtain three feature sets of different scales: a contact probability distribution and genome distance feature set, a Bin contact probability feature set, and a small domain contact probability feature set; The contact probability distribution and genome distance feature set is specifically: The interaction of two fragments is regarded as a contact, and the contact is distributed in multiple intervals divided according to the linear distance of the two fragments in the contact; The probability of contact count in each interval is calculated to obtain the contact probability distribution and genome distance feature set of the cell; The Bin contact probability feature set is specifically: Based on the chromosome contact matrix, the contact probability of each Bin on the chromosome is calculated; by calculating the Bin contact probability of all chromosomes, the Bin contact probability feature set of the cell is generated; The small domain contact probability feature set is specifically: The small domain is defined as a triangular region with a target Bin as the center and first-order linear adjacent Bins as the boundary; Based on the chromosome contact matrix, the contact probability of each small domain is calculated; By calculating the small domain contact probability of all chromosomes, the small domain contact probability feature set of the cell is generated; The three extracted feature sets are input into the trained fusion prediction model for convolution and feature fusion to generate and output the final cell cycle prediction result.

2. The cell cycle prediction method based on single-cell Hi-C data according to claim 1, wherein, The preprocessing is specifically: Based on the read pair site mapping file in the single cell Hi-C data, the interaction information based on fragment number in the single cell Hi-C data is converted into interaction information based on fragment position coordinates; All chromosomes in the single cell are divided into multiple fragments according to a preset resolution, and each fragment is called a Bin; According to the interaction information identified based on the fragment coordinates, the interaction information is mapped to the corresponding Bin to generate a corresponding chromosome contact matrix for each chromosome of the single cell.

3. The cell cycle prediction method based on single-cell Hi-C data according to claim 1, wherein, The fusion prediction model comprises a convolution module and a feature fusion module; The convolution module generates a one-dimensional feature vector for each feature set through a one-dimensional convolution layer, a batch normalization layer, a max pooling layer, and a Dropout layer; The feature fusion module fuses the three one-dimensional feature vectors into a final cell feature, and scores the cell feature in stages, and the stage with the highest score is the final cell cycle prediction result.

4. The cell cycle prediction method based on single-cell Hi-C data according to claim 1, wherein, The training of the fusion prediction model adopts a 5-fold cross-validation method, and a preset proportion of training set data is segmented into a validation set every fold, and the loss on the validation set does not decrease in a continuous preset number of rounds, and then the model training is terminated.

5. A cell cycle prediction system based on single-cell Hi-C data, characterized in that, The method comprises a data preprocessing module, a feature extraction module, and a cycle prediction module: The data preprocessing module is configured to preprocess the Hi-C data of a single cell to be predicted to obtain a chromosome contact matrix; The feature extraction module is configured to extract single-cell chromosome three-dimensional structure features based on interaction information of a single cell to be predicted and a chromosome contact matrix, to obtain three feature sets of different scales: a contact probability distribution and genome distance feature set, a Bin contact probability feature set, and a small-domain contact probability feature set. The contact probability distribution and genome distance feature set is specifically: Interaction of two fragments is taken as a contact, and the contact is distributed into multiple intervals divided according to linear distances of two fragments in the contact; The probability of contact count in each interval is calculated to obtain the contact probability distribution and genome distance feature set of the cell; The Bin contact probability feature set is specifically: The contact probability of each Bin on the chromosome is calculated based on the chromosome contact matrix; the Bin contact probability feature set of the cell is generated by calculating Bin contact probabilities of all chromosomes; The small-domain contact probability feature set is specifically: The small domain is defined as a triangular region with a target Bin as a center and first-order linear adjacent Bins as boundaries; The contact probability of each small domain is calculated based on the chromosome contact matrix; the small-domain contact probability feature set of the cell is generated by calculating small-domain contact probabilities of all chromosomes; The cycle prediction module is configured to input the three extracted feature sets into a trained fusion prediction model, perform convolution and feature fusion, and generate and output a final cell cycle prediction result. 6.An electronic device, comprising: a memory for non-transitory storage of computer readable instructions; and a processor for running the computer readable instructions, wherein the computer readable instructions, when executed by the processor, perform the method of any one of claims 1-4.

7. A storage medium characterized by non-transitorily storing computer readable instructions, wherein, When the non-transitory computer readable instructions are executed by a computer, instructions for performing the method of any one of claims 1-4.

Images