Intelligent chromatin interaction prediction method and system based on nanopore sequencing
By constructing positive and negative samples of chromatin loops, extracting methylation and sequence characteristics, combining XGBoost and convolutional neural network models, the high cost and low efficiency problems of chromatin interaction prediction in the existing technology are solved, and efficient and accurate chromatin interaction prediction is achieved. It is suitable for interactions between different individuals and long distances, and has clinical application potential.
Patent Information
- Application Number
- CN202510638376.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-19
- Publication Date
- 2025-08-15
AI Technical Summary
The existing chromatin interaction prediction methods are expensive, inefficient, and have limited scope of application, so they cannot directly predict from third-generation sequencing data.
The positive and negative samples of chromatin loop were constructed, the methylation characteristics and sequence characteristics were extracted, and the XGBoost and convolutional neural network model were combined with nanopore sequencing data were used to predict the final probability.
Efficient and accurate prediction of chromatin interactions is suitable for different individuals, considering genetic differences and long-distance interactions, reducing costs, and is suitable for large-scale clinical testing.
Smart Images
Figure CN120496629A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of bioinformatics, and in particular to a method and system for intelligent prediction of chromatin interactions based on nanopore sequencing. Background Art
[0002] DNA is not only the blueprint of life but also forms the three-dimensional structure that controls cellular function and gene regulation. Interactions within chromatin are fundamental to organizing the spatial layout of the genome, enabling close contact between regulatory elements such as enhancers, promoters, and insulators and target genes. These interactions help precisely control gene transcriptional activity, influence gene expression patterns, and maintain a stable state within the cell. Understanding the interaction mechanisms of chromatin is crucial for understanding genome organization and its impact on development, disease, and cellular function.
[0003] In recent years, with the development of chromatin conformation capture technologies (such as 3C), especially the application of advanced methods such as Hi-C, ChIA-PET, and HiChIP / PLAC-Seq, we can now systematically study chromatin interactions on a genome-wide scale. These technologies provide unprecedented three-dimensional genomic details, but their high cost and complex experimental procedures limit their widespread application in various cell types and clinical settings. Therefore, it is particularly urgent to develop computational tools that can efficiently and economically predict chromatin interactions.
[0004] Recent advances in deep learning have brought new opportunities for predicting chromatin interactions. For example, methods such as DeepSea and DeepBind have demonstrated the ability to predict transcription factor binding sites from DNA sequences. The subsequent emergence of the ChINN model based on convolutional neural networks (CNNs) further demonstrated that interactions between open chromatin regions can be predicted based on DNA sequence alone. In addition, the CLNN loop is a method that uses a bidirectional long short-term memory network (LSTM) to predict chromatin loops and CTCF binding site pairs in different cell lines by integrating multiple sequence features. DeepCTCFLoop can distinguish CTCF motif pairs that can form chromatin loops from motifs that cannot form loops based solely on the DNA sequence of the CTCF motif and its flanking regions. Although these models can predict chromatin interactions from DNA sequences, they also have certain limitations: ChINN uses a single reference genome for prediction and fails to consider genetic differences between individuals, which limits its applicability to different samples. At the same time, it relies on distance features and has difficulty accurately predicting long-range interactions. CLNN loops and DeepCTCFLoop mainly focus on predicting CTCF-related loop structures, and their ability to predict non-CTCF-mediated loop structures is still unclear. In addition, the application effect of these methods on third-generation sequencing data has not been verified.
[0005] Research has shown that DNA methylation plays a crucial role in shaping three-dimensional chromatin structure and regulating gene expression. Incorporating precise DNA sequence and methylation information into predictive models can significantly improve the accuracy of chromatin interaction predictions, thereby providing deeper insights into the relationship between epigenetics and genomic organization. Third-generation sequencing (TGS) technologies, such as PacBio's single-molecule real-time (SMRT) sequencing and Oxford Nanopore Technology (ONT), have revolutionized this field. These platforms not only provide high-resolution DNA sequence and methylation data but also avoid the bias and errors associated with PCR amplification. The long reads generated by TGS also help capture complex genomic interactions. Furthermore, PacBio and ONT technologies, powered by advanced algorithms, can detect DNA methylation directly from raw sequencing signals, enabling precise methylation analysis and providing new insights into epigenetic regulation. Although TGS has significantly advanced data generation for sequence and methylation signatures, current computational methods are still unable to directly predict chromatin interactions from TGS data. This gap requires us to develop innovative methods that fully utilize the dual advantages provided by TGS - high-resolution sequence and methylation information to fill the gap between raw sequencing data and three-dimensional genome organization. Summary of the Invention
[0006] (1) Technical issues to be resolved
[0007] Based on the above problems, the present invention provides a method and system for intelligent prediction of chromatin interactions based on nanopore sequencing, which solves the problems that existing chromatin interaction prediction methods are expensive, inefficient, have limited applicability, and cannot be predicted directly from TGS data.
[0008] (2) Technical solution
[0009] Based on the above technical problems, the present invention provides a method for intelligent prediction of chromatin interactions based on nanopore sequencing, comprising:
[0010] S1. Construct positive and negative samples of chromatin loops: Identify positive samples of chromatin loops from .hic data and record them as loops pos ; Anchor points are collected from the whole genome through a sliding window, and the anchor points are randomly grouped into anchor point pairs. The anchor point pairs are sampled according to the distance distribution of the chromatin loop positive samples to generate chromatin loop negative samples, which are recorded as loop neg ;
[0011] S2. Generate methylation signatures: Based on WGBS, extract the number of CpG sites with an average methylation probability ≥ 0.5 from the positive and negative samples of the chromatin loop, and calculate the Pearson correlation coefficient based on each anchor point pair to obtain the methylation signature of the positive and negative samples of the chromatin loop, denoted as methyl;
[0012] S3. Generate a submatrix set: Extract DNA sequences from the DNA sequence data based on the range of anchor points in the positive and negative samples of chromatin loops, obtain a sequence matrix after encoding, and split the sequence matrix and the reverse complementary sequence matrix separately to obtain a submatrix set;
[0013] S4. Generate preliminary probabilities based on methylation features: Input the methylation features of the positive and negative samples of the chromatin loop into the methylation module for preliminary prediction and output preliminary probabilities, recorded as out1;
[0014] S5. Extracting sequence features based on the sub-matrix set: Inputting the sub-matrix set into the sequence module respectively, extracting sequence features, and concatenating the four sets of sequence features into a feature, which is recorded as out2;
[0015] S6. Generate the final probability based on the methylation features, preliminary probability, sequence features, and distance features: Calculate the distance between the centers of the left and right anchor points of each anchor point pair to obtain the distance feature, recorded as distance; then input the concatenated features of methyl, out1, out2, and distance into the regressor in the sequence module, and output the final probability, which represents the probability that the final predicted anchor point pair will form a chromatin loop.
[0016] Furthermore, in S1, the sliding window is a window with a length of 5000 bp and a sliding step of 500 bp, so that there is a gap of at least 500 bp between the collected anchor point and the anchor point of the chromatin loop positive sample.
[0017] Furthermore, in S1, the anchor point pairs are sampled according to the distance distribution of the chromatin loop positive sample, including: counting the distribution number of the distance of the chromatin loop positive sample in the three intervals of 0-100000, 100000-1000000, and above 1000000, sampling the anchor point pairs so that the number of anchor point pairs in the three intervals is the same as the number of the chromatin loop positive samples, and generating a chromatin loop negative sample.
[0018] Furthermore, the S2 includes:
[0019] S21. Divide the anchor point pairs in the positive and negative samples of the chromatin loop into multiple intervals according to the left anchor point and the right anchor point respectively. Then, based on the WGBS data, count the number of CpG sites with an average methylation probability ≥ 0.5 in each interval to obtain the methylation signature. The methylation signatures obtained from the left anchor point and the right anchor point are respectively recorded as methyl L methyl R ;
[0020] S22, for each anchor point pair corresponding to the methyl L and methyl R Calculate the Pearson correlation coefficient, denoted as p;
[0021] S23, the methyl L , methyl R and p to obtain the methylation characteristics of the positive and negative samples of the chromatin loop, which is recorded as methyl.
[0022] Furthermore, the S3 includes:
[0023] S31. Extract DNA sequences from the DNA sequence data according to the range of anchor points in the positive and negative samples of chromatin loops, and encode them into 4-dimensional matrices. The matrices obtained from the left anchor point and the right anchor point are denoted as LF and RF, respectively.
[0024] S32, based on the pairing principle of AT and GC, generate complementary sequences of LF and RF, and then invert the complementary sequences to generate reverse complementary sequence matrices, which are recorded as LRC and RRC respectively;
[0025] S33. Divide the matrices LF, RF, LRC, and RRC into several sub-matrices respectively to obtain four sub-matrix sets.
[0026] Furthermore, the S33 includes: dividing LF, RF, LRC and RRC into several sub-matrices with a length of 1000 bp respectively, with each sub-matrix overlapping by 500 bp, and the obtained four groups of sub-matrices are respectively recorded as LFs, RFs, LRCs and RRCs.
[0027] Furthermore, in S4, the methylation module includes an XGBoost model.
[0028] Furthermore, in S5, the sequence module includes a convolutional neural network, which includes a first convolutional layer, a second convolutional layer, and a third convolutional layer. The first convolutional layer and the second convolutional layer both include sequentially connected conv1 convolution, ReLU activation function, maximum pooling layer and random dropout layer, and the third convolutional layer includes sequentially connected conv1 convolution, random dropout layer, weighted sum layer and Tanh activation function, and finally outputs a 128-dimensional output vector.
[0029] Furthermore, in S5, the weighted sum layer defines a weight matrix with a shape of (128, 53). Assuming that x is the output after processing by the maximum pooling layer and the dropout layer, and the shape is (batch_size, 128, seq_len), then the output of the weighted sum layer is described as:
[0030]
[0031] Among them, x i is the i-th element in the convolution output, w i is the weight coefficient in the weighted summation layer.
[0032] The present invention also discloses a chromatin interaction intelligent prediction system based on nanopore sequencing, which runs the chromatin interaction intelligent prediction method based on nanopore sequencing, including a methylation module and a sequence module. The methylation module runs S2 and S4, and the sequence module runs S3, S5, and S6.
[0033] (3) Beneficial effects
[0034] The above technical solution of the present invention has the following advantages:
[0035] (1) The present invention obtains positive samples of chromatin loops from Hi-C data, generates negative samples of chromatin loops from whole-genome sampling, extracts methylation features from the constructed positive and negative samples of chromatin loops, obtains preliminary probabilities through preliminary prediction of the methylation features, and extracts sequence features through a convolutional neural network. The final probability of predicted anchor pairs forming chromatin loops is generated based on the methylation features, preliminary probabilities, sequence features, and distance features. This allows for direct identification of chromatin interactions from nanopore sequencing data, filling the research gap in predicting chromatin interactions from third-generation sequencing (TGS) data.
[0036] (2) The present invention can accurately predict chromatin interactions in Hi-C data, and has a wider applicability than existing technologies that can only predict CTCF-mediated interactions. It can also detect chromatin loops based on nanopore sequencing data from different individuals, taking into account genetic differences between individuals, making it applicable to different samples. It also takes into account distance characteristics and can accurately predict long-range interactions, showing strong adaptability. Therefore, the present invention has the potential to be applied to large-scale clinical testing.
[0037] (3) The present invention integrates DNA sequence and methylation information in nanopore sequencing data without the need for high-throughput experimental data such as Hi-C, and efficiently and accurately predicts the probability of chromatin loop formation. It can also be run on a general computing server or personal PC, saving the high experimental cost and complex process of implementing the chromatin interaction capture method, with very low recognition cost and higher recognition efficiency. BRIEF DESCRIPTION OF THE DRAWINGS
[0038] The features and advantages of the present invention will be more clearly understood by referring to the accompanying drawings, which are schematic and should not be construed as limiting the present invention in any way. In the accompanying drawings:
[0039] Figure 1 Schematic diagram of the process of the intelligent prediction method of chromatin interactions based on nanopore sequencing according to an embodiment of the present invention. DETAILED DESCRIPTION
[0040] The following embodiments of the present invention are described in further detail with reference to the accompanying drawings and examples. The following examples are used to illustrate the present invention but are not intended to limit the scope of the present invention.
[0041] The present invention provides an intelligent prediction method for chromatin interactions based on nanopore sequencing. Figure 1 As shown, specifically including:
[0042] S1. Construct positive and negative samples of chromatin loops: Identify positive samples of chromatin loops from .hic data and record them as loops pos; Anchor points are collected from the whole genome through a sliding window, and the anchor points are randomly grouped into anchor point pairs. The anchor point pairs are sampled according to the distance distribution of the chromatin loop positive samples to generate chromatin loop negative samples, which are recorded as loop neg ; Specifically including the following steps:
[0043] S11. Identify chromatin loop positive samples from .hic data, denoted as loop pos ;
[0044] Use the HiCCUPS tool to identify loop chromatin structures from .hic data. The command is java -Xmx20g -jar juicer_tools.jar hiccups -r 5000 -k MboI -t 40 -p 10 -i 5000 -o sample1 / sample1.hic, where -r represents the resolution parameter, which needs to be set to 5000bp. The identified chromatin loop positive samples are recorded as loops. pos ;
[0045] S12, collect anchor points from the whole genome through a sliding window;
[0046] A window of 5000 bp in length and a sliding step of 500 bp is used to slide across the entire genome to generate random anchors. It is required that there is at least a 500 bp gap between the random anchor and the positive sample anchor.
[0047] S13, randomly grouping the anchor points into anchor point pairs, sampling the anchor point pairs according to the distance distribution of the chromatin loop positive samples, and generating chromatin loop negative samples, recorded as loop neg ;
[0048] The anchor points generated by sliding are randomly combined into anchor pairs, and the number of chromatin loop positive samples with distances between 0-100,000, 100,000-1,000,000, and above 1,000,000 is counted. The randomly composed anchor pairs are sampled to ensure that the number in each interval is consistent with the number of chromatin loop positive samples. The generated chromatin loop negative samples are recorded as loops. neg , and finally generate 1:1 loop chromatin positive and negative samples.
[0049] S2. Generate methylation signatures: Based on WGBS, extract the number of CpG sites with an average methylation probability ≥ 0.5 from the positive and negative samples of the chromatin loop, and calculate the Pearson correlation coefficient based on each anchor point pair to obtain the methylation signatures of the positive and negative samples of the chromatin loop, denoted as methyl. This specifically includes the following steps:
[0050] S21. Divide the anchor point pairs in the positive and negative samples of the chromatin loop into multiple intervals according to the left anchor point and the right anchor point respectively. Then, based on the WGBS data, count the number of CpG sites with an average methylation probability ≥ 0.5 in each interval to obtain the methylation signature. The methylation signatures obtained from the left anchor point and the right anchor point are respectively recorded as methyl L methyl R ;
[0051] Input loop chromatin loop positive and negative samples and WGBS (Whole Genome Bisulfite Sequencing) data, and convert loop pos and loop neg The left and right anchors in the anchor pair are divided into 10 intervals respectively. Then, based on the WGBS data, the number of CpG sites with an average methylation probability ≥ 0.5 in each interval is counted (regions with at least 200 bases and a GC content of more than 50% are called CpG sites or CpG islands). The number is counted for each of the 10 intervals to obtain the 10-dimensional methylation feature of the left anchor, which is recorded as methyl L , the 10-dimensional methylation feature of the right anchor is recorded as methyl R , the two anchors of a loop can generate two 10-dimensional features;
[0052] S22, for each anchor point pair corresponding to the methyl L and methyl R Calculate the Pearson correlation coefficient, denoted as p;
[0053] For each anchor pair, methyl L and methyl R Calculate the Pearson correlation coefficient. The calculated Pearson correlation coefficient is recorded as p. The formula is as follows:
[0054]
[0055] Among them, Xi and Yi are the two axis coordinates of the sample point, and are the standard scores of the corresponding coordinates, and n is the total number of sample points;
[0056] S23, the methyl L , methyl R and p to obtain the methylation characteristics of the positive and negative samples of the chromatin loop, which is recorded as methyl;
[0057] MethylL , methyl R and p, so that a chromatin loop positive and negative sample loop can obtain a 21-dimensional methylation feature, recorded as methyl.
[0058] S3. Generate a submatrix set: Extract DNA sequences from the DNA sequence data based on the range of anchor points in the positive and negative samples of the chromatin loop, obtain a sequence matrix through encoding, and split the sequence matrix and the reverse complementary sequence matrix separately to obtain a submatrix set; specifically, the following steps are included:
[0059] S31. Extract DNA sequences from the DNA sequence data according to the range of anchor points in the positive and negative samples of chromatin loops, and encode them into 4-dimensional matrices. The matrices obtained from the left anchor point and the right anchor point are denoted as LF and RF, respectively.
[0060] Input loop pos and loop neg As well as DNA sequence data, according to the range of the anchor point, the DNA sequences of the left and right anchor points are taken out, and then they are encoded into 4-dimensional matrices using one-hot encoding. Each dimension corresponds to a base ATGC, A represents adenine, T represents thymine, G represents guanine, and C represents cytosine. The matrices after encoding the left and right anchor points are denoted as LF and RF respectively;
[0061] S32, based on the pairing principle of AT and GC, generate complementary sequences of LF and RF, and then invert the complementary sequences to generate reverse complementary sequence matrices, which are recorded as LRC and RRC respectively;
[0062] According to the pairing principle of AT and GC, the complementary sequences of LF and RF are generated, and then the complementary sequences are reversed to generate the reverse complementary sequence matrices, which are recorded as LRC and RRC respectively;
[0063] S33, dividing the matrices LF, RF, LRC, and RRC into several sub-matrices respectively, to obtain four sets of sub-matrices;
[0064] LF, RF, LRC, and RRC are divided into several sub-matrices of 1000 bp in length, with 500 bp overlap between each sub-matrix. The resulting four sub-matrix sets are denoted as LFs, RFs, LRCs, and RRCs respectively; each sub-matrix set is subsequently input into the sequence module for feature extraction.
[0065] S4. Generate preliminary probabilities based on methylation features: Input the methylation features of the positive and negative samples of chromatin loops into the methylation module for preliminary prediction and output a preliminary probability, denoted as out1, which represents the possibility of a given anchor point pair forming a chromatin loop;
[0066] To make a preliminary prediction of the 21-dimensional methylation features of anchor pairs, an XGBoost-based methylation module was used. The 21-dimensional methylation features of the anchor pairs (methyl) were input into the methylation module, which is constructed using an XGBoost model. The model outputs a probability value between 0 and 1 for the formation of a chromatin loop, representing the preliminary predicted probability of chromatin loop formation, denoted as out1. This module optimizes the XGBoost model using the following parameter configurations to ensure accurate and efficient predictions:
[0067] (1) Objective function: binary:logistic, used for binary classification tasks, outputs the probability value of forming a chromatin loop.
[0068] (2) Maximum tree depth (Max Depth): max_depth, controls the maximum depth of each tree to prevent overfitting.
[0069] (3) Learning rate (Eta): 0.1, adjusts the learning step size of each iteration to balance the convergence speed and model performance.
[0070] (4) Number of threads (Nthread): threads, specifies the number of threads for parallel computing to improve training efficiency.
[0071] (5) Evaluation metrics (EvalMetric): including aucpr (area under the average precision-recall curve), map (mean average precision) and logloss (logarithmic loss), which are used to monitor model performance.
[0072] (6) L2 regularization term weight (Lambda): 1.0, which increases the complexity penalty of the model to prevent overfitting.
[0073] (7) Subsample ratio: 0.8, randomly select 80% of the data to train each tree to introduce diversity.
[0074] (8) Column sampling ratio (Colsample_bytree): 0.5, randomly select 50% of the features for training each tree to further enhance the generalization ability of the model.
[0075] Specifically, the 21-dimensional methylation features are fed into the aforementioned XGBoost model. After multiple rounds of iterative training, the model outputs a probability value between 0 and 1, indicating the likelihood of a given anchor point pair forming a chromatin loop. This probability value serves as an important basis for subsequent analysis and decision-making.
[0076] S5. Extracting sequence features based on the sub-matrix set: Inputting the sub-matrix set into the sequence module respectively, extracting sequence features, and concatenating the four sets of sequence features into a feature, which is recorded as out2;
[0077] The four submatrix sets LFs, RFs, LRCs, and RRCs generated in S3 are respectively input into a sequence module composed of a convolutional neural network (CNN) to extract sequence features; the convolutional neural network includes the first convolution layer, the second convolution layer, and the third convolution layer. The first convolution layer and the second convolution layer both include sequentially connected conv1 convolution, ReLU activation function, maxpool maximum pooling layer, and dropout random inactivation layer. The third convolution layer includes sequentially connected conv1 convolution, dropout random inactivation layer, Weighted_sum weighted summation layer, and Tanh activation function. The final output is a 128-dimensional output vector. The four sets of output vectors are concatenated to form a 512-dimensional feature, recorded as out2. This sequence module is based on a customized deep learning architecture. The specific implementation of the model architecture is as follows:
[0078] First convolutional layer: Use nn.Conv1d for one-dimensional convolution, with 4 input channels, 128 output channels, and a convolution kernel size of 8.
[0079] Second convolutional layer: Continue to use nn.Conv1d, with 128 input channels, 256 output channels, and a convolution kernel size of 8.
[0080] The third convolutional layer uses nn.Conv1d again, with 256 input channels, 128 output channels, and a kernel size of 8.
[0081] After the first and second convolutional layers, the ReLU activation function is used to introduce nonlinearity, followed by a MaxPool layer with a pooling window size of 4. This means that each pooling operation takes four consecutive elements from each channel of the input tensor and selects the maximum of these four elements as the output. The stride is 4, indicating that the pooling window moves four positions at a time, reducing the size of the feature map and extracting important features. A dropout layer with a probability of 0.5 is connected after the MaxPool layer. This means that during the forward propagation, each neuron has a 50% probability of being dropped and a 50% probability of remaining unchanged, thus forcing the network to learn more robust feature representations.
[0082] To further enhance the expressiveness of the model, a trainable weighted sum layer Weighted_sum is introduced in the third convolutional layer. This layer defines a weight matrix with a shape of (128, 53) through torch.nn.Parameter. Assuming that x is the output after processing through the above convolutional layer and pooling layer, with a shape of (batch_size, 128, seq_len), the output of weighted_sum can be described by the following expression:
[0083]
[0084] where x i is the i-th element in the convolution output, w i is the weighting coefficient in weighted_sum. During forward propagation, the feature map is multiplied by this weight matrix and summed along the specified dimension to produce a final 128-dimensional output vector. This operation helps the model better focus on important features, thereby improving its expressiveness and performance. A Tanh activation function is also connected after the weighted sum layer to enhance the model's nonlinear representation capabilities.
[0085] After each sub-matrix is processed by the convolutional neural network, a 128-dimensional output vector is obtained. The output vectors of the four sub-matrices are concatenated together to form a 512-dimensional feature vector for subsequent analysis and prediction.
[0086] S6. Generate the final probability based on the methylation features, preliminary probability, sequence features, and distance features: Calculate the distance between the centers of the left and right anchor points of each anchor point pair to obtain the distance feature, recorded as distance; then input the concatenated features of methyl, out1, out2, and distance into the regressor in the sequence module, and output the final probability, which represents the probability that the final predicted anchor point pair will form a chromatin loop.
[0087] The 22-dimensional input and output features of the methylation module, the 512-dimensional features of the sequence module, and the distance features of the anchorpair are concatenated into 535-dimensional features, which are then input into the regressor composed of XGBoost. The regressor will output a probability between 0 and 1, indicating the final predicted probability of loop formation.
[0088] For each anchor pair, the distance between the centers of the two anchor points is calculated, recorded as distance. The methyl, out1, out2 and distance features are then concatenated into a 535-dimensional feature. This feature is then input into the regressor constructed by XGBoost in the sequence module. The regressor outputs a probability between 0 and 1, indicating the probability that the final predicted anchor pair will form a chromosome ring.
[0089] For the above examples, to verify the effectiveness of the present invention, we used ChINN, CLNN-Loop, and DeepCTCFLoop, along with the present invention, to predict chromatin interactions using DNA sequence and methylation information obtained from nanopore sequencing data from four cells. The performance of the method was then measured using the area under the Precision-Recall curve (AUPRC). This is a metric that can be used to evaluate the performance of a regression model. Generally, an AUPRC value closer to 1 indicates better model performance; an AUPRC value closer to 0 indicates worse model performance.
[0090] Table 1. Performance comparison with DNA sequence-based chromatin interaction prediction methods
[0091]
[0092] As shown in Table 1, across 16 test experiments across four datasets, the methods proposed in this paper for predicting chromatin interactions using nanopore sequencing data all achieved the highest AUPRCs, with all achieving AUPRCs of 0.88 or higher. To further validate the broad applicability of this invention, we trained each of the four methods on one cell line and performed predictions on the remaining cell lines, testing their cross-cell line prediction performance. The results are shown in Table 2.
[0093] Table 2. Comparison of performance across cell lines with DNA sequence-based chromatin interaction prediction methods
[0094]
[0095]
[0096] As can be seen from Table 2, the present invention achieved the highest performance in all 12 cross-cell line prediction experiments, indicating that the present invention has fully learned the universal characteristics of chromatin interactions and can achieve good results even when predicting unprecedented cell lines, demonstrating that the present invention has the potential to be widely used in large-scale clinical testing.
[0097] The prediction method of this embodiment can also be implemented by a prediction system. The prediction system consists of two parts: a methylation module and a sequence module. The methylation module runs S2 and S4, and the sequence module runs S3, S5, and S6.
[0098] Finally, it should be noted that the above-mentioned prediction method can be converted into software program instructions, which can be implemented by using a prediction system including a processor and a memory, or by computer instructions stored in a non-transitory computer-readable storage medium. The above-mentioned integrated unit implemented in the form of a software functional unit can be stored in a computer-readable storage medium. The above-mentioned software functional unit is stored in a storage medium, including a number of instructions for enabling a computer device (which can be a personal computer, a server, or a network device, etc.) or a processor to perform some steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), disk or optical disk, and other media that can store program code.
[0099] In summary, the above-mentioned intelligent prediction method for chromatin interactions based on nanopore sequencing has the following beneficial effects:
[0100] (1) The present invention obtains positive samples of chromatin loops from Hi-C data, generates negative samples of chromatin loops from whole genome sampling, extracts methylation features from the generated positive and negative samples of chromatin loops, obtains preliminary probabilities through preliminary prediction of the methylation features, and extracts sequence features through a convolutional neural network. The final probability of predicted anchor points forming chromatin loops is generated based on the methylation features, preliminary probabilities, sequence features, and distance features. This allows for direct identification of chromatin interactions from nanopore sequencing data, filling the research gap in predicting chromatin interactions from third-generation sequencing (TGS) data.
[0101] (2) The present invention can accurately predict chromatin interactions in Hi-C data, and has a wider applicability than existing technologies that can only predict CTCF-mediated interactions. It can also detect chromatin loops based on nanopore sequencing data from different individuals, taking into account genetic differences between individuals, making it applicable to different samples. It also takes into account distance characteristics and can accurately predict long-range interactions, showing strong adaptability. Therefore, the present invention has the potential to be applied to large-scale clinical testing.
[0102] (3) The present invention integrates DNA sequence and methylation information in nanopore sequencing data without the need for high-throughput experimental data such as Hi-C, and efficiently and accurately predicts the probability of chromatin loop formation. It can also be run on a general computing server or personal PC, saving the high experimental cost and complex process of implementing the chromatin interaction capture method, with very low recognition cost and higher recognition efficiency.
[0103] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit the present invention. Although the embodiments of the present invention have been described in conjunction with the accompanying drawings, those skilled in the art may make various modifications and variations without departing from the spirit and scope of the present invention, and such modifications and variations shall fall within the scope defined by the appended claims.
Claims
1. A method for intelligent prediction of chromatin interactions based on nanopore sequencing, characterized in that: include: S1. Construct positive and negative samples of chromatin loops: Identify positive samples of chromatin loops from .hic data and record them as loops pos ; Anchor points are collected from the whole genome through a sliding window, and the anchor points are randomly grouped into anchor point pairs. The anchor point pairs are sampled according to the distance distribution of the chromatin loop positive samples to generate chromatin loop negative samples, which are recorded as loop neg ; S2. Generate methylation signatures: Based on WGBS, extract the number of CpG sites with an average methylation probability ≥ 0.5 from the positive and negative samples of the chromatin loop, and calculate the Pearson correlation coefficient based on each anchor point pair to obtain the methylation signature of the positive and negative samples of the chromatin loop, denoted as methyl; S3. Generate a submatrix set: Extract DNA sequences from the DNA sequence data based on the range of anchor points in the positive and negative samples of chromatin loops, obtain a sequence matrix after encoding, and split the sequence matrix and the reverse complementary sequence matrix separately to obtain a submatrix set; S4. Generate preliminary probabilities based on methylation features: Input the methylation features of the positive and negative samples of the chromatin loop into the methylation module for preliminary prediction and output preliminary probabilities, recorded as out1; S5. Extracting sequence features based on the sub-matrix set: Inputting the sub-matrix set into the sequence module respectively, extracting sequence features, and concatenating the four sets of sequence features into a feature, which is recorded as out2; S6. Generate the final probability based on the methylation features, preliminary probability, sequence features, and distance features: Calculate the distance between the centers of the left and right anchor points of each anchor point pair to obtain the distance feature, recorded as distance; then input the concatenated features of methyl, out1, out2, and distance into the regressor in the sequence module, and output the final probability, which represents the probability that the final predicted anchor point pair will form a chromatin loop.
2. The method for intelligent prediction of chromatin interactions based on nanopore sequencing according to claim 1, characterized in that In S1, the sliding window is a window with a length of 5000 bp and a sliding step of 500 bp, so that there is a gap of at least 500 bp between the collected anchor point and the anchor point of the chromatin loop positive sample.
3. The method for intelligent prediction of chromatin interactions based on nanopore sequencing according to claim 1, characterized in that In S1, the anchor point pairs are sampled according to the distance distribution of the chromatin loop positive sample, including: counting the distribution number of the distance of the chromatin loop positive sample in three intervals of 0-100000, 100000-1000000, and above 1000000, sampling the anchor point pairs so that the number of anchor point pairs in the three intervals is the same as the number of the chromatin loop positive sample, and generating a chromatin loop negative sample.
4. The method for intelligent prediction of chromatin interactions based on nanopore sequencing according to claim 1, characterized in that The S2 includes: S21. Divide the anchor point pairs in the positive and negative samples of the chromatin loop into multiple intervals according to the left anchor point and the right anchor point respectively. Then, based on the WGBS data, count the number of CpG sites with an average methylation probability ≥ 0.5 in each interval to obtain the methylation signature. The methylation signatures obtained from the left anchor point and the right anchor point are respectively recorded as methyl L methyl R ; S22, for each anchor point pair corresponding to the methyl L and methyl R Calculate the Pearson correlation coefficient, denoted as p; S23, the methyl L , methyl R and p to obtain the methylation characteristics of the positive and negative samples of the chromatin loop, which is recorded as methyl.
5. The method for intelligent prediction of chromatin interactions based on nanopore sequencing according to claim 1, characterized in that The S3 includes: S31. Extract DNA sequences from the DNA sequence data according to the range of anchor points in the positive and negative samples of chromatin loops, and encode them into 4-dimensional matrices. The matrices obtained from the left anchor point and the right anchor point are denoted as LF and RF, respectively. S32, based on the pairing principle of AT and GC, generate complementary sequences of LF and RF, and then invert the complementary sequences to generate reverse complementary sequence matrices, which are recorded as LRC and RRC respectively; S33. Divide the matrices LF, RF, LRC, and RRC into several sub-matrices respectively to obtain four sub-matrix sets.
6. The method for intelligent prediction of chromatin interactions based on nanopore sequencing according to claim 5, characterized in that The S33 includes: dividing LF, RF, LRC and RRC into several sub-matrices with a length of 1000 bp respectively, with each sub-matrix overlapping by 500 bp, and the obtained four groups of sub-matrices are respectively recorded as LFs, RFs, LRCs and RRCs.
7. The method for intelligent prediction of chromatin interactions based on nanopore sequencing according to claim 1, characterized in that: In S4, the methylation module includes an XGBoost model.
8. The method for intelligent prediction of chromatin interactions based on nanopore sequencing according to claim 1, characterized in that: In S5, the sequence module includes a convolutional neural network, which includes a first convolutional layer, a second convolutional layer, and a third convolutional layer. The first convolutional layer and the second convolutional layer both include sequentially connected conv1 convolution, ReLU activation function, maximum pooling layer and random dropout layer. The third convolutional layer includes sequentially connected conv1 convolution, random dropout layer, weighted sum layer and Tanh activation function, and the final output is a 128-dimensional output vector.
9. The method for intelligent prediction of chromatin interactions based on nanopore sequencing according to claim 8, characterized in that: In S5, the weighted sum layer defines a weight matrix with a shape of (128, 53). Assuming that x is the output after processing by the maximum pooling layer and the dropout layer, and the shape is (batch_size, 128, seq_len), the output of the weighted sum layer is described as: Among them, x i is the i-th element in the convolution output, w i is the weight coefficient in the weighted summation layer.
10. A chromatin interaction intelligent prediction system based on nanopore sequencing, characterized in that: Running the nanopore sequencing-based intelligent prediction method for chromatin interactions according to any one of claims 1 to 9, comprising a methylation module and a sequence module, wherein the methylation module runs S2 and S4, and the sequence module runs S3, S5, and S6.