Method for identifying chromatin structural features from Hi-C matrix, non-transitory computer readable medium storing a program for identifying chromatin structural features from Hi-C matrix

通过计算Hi-C矩阵的相关矩阵和主成分分析,识别染色质结构特征,解决了现有技术中难以识别多尺度染色质结构和区分细胞类型的问题,实现了对癌症的准确诊断和治疗。

CN116981779BActive Publication Date: 2024-11-08CHROMATIN (BEIJING) TECH CO LTD
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202280000359.1
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2022-02-08
Publication Date
2024-11-08
Estimated Expiration
2042-02-08

AI Technical Summary

Technical Problem

The existing Hi-C matrix processing methods are difficult to accurately identify multi-scale chromatin structural characteristics, especially in the case of poor data quality or high noise, and cannot effectively distinguish the structural differences between normal and abnormal cells, affecting cancer diagnosis and treatment. .

Method used

By calculating the correlation matrix of the Hi-C matrix, the structural feature vector is derived, and the structural features of chromatin are identified using principal component analysis, and the cell type map is used to distinguish different cell types, and the differences between normal and abnormal cells are identified.

Benefits of technology

The more reliable and effective identification of chromatin structural differences in the Hi-C matrix is ​​achieved, providing potential targets for cancer diagnosis and treatment, and improving the accuracy of early screening and treatment of cancer.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN116981779B_ABST
    Figure CN116981779B_ABST
Patent Text Reader

Abstract

The present invention relates to a method for identifying chromatin structural features from a Hi-C matrix, a non-transient computer-readable medium storing a program for identifying chromatin structural features from a Hi-C matrix, and a method for diagnosing and treating a medical condition or disease. The method for identifying chromatin structural features from a Hi-C matrix comprises performing correlation processing on the Hi-C matrix to calculate a correlation matrix; calculating a structural feature vector based on the correlation matrix; calculating a principal component score matrix based on the structural feature vector; and identifying at least one chromatin structural feature in the principal component score matrix.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present application relates to methods for identifying chromatin structural features from Hi-C matrices, non-transitory computer-readable media storing programs for identifying chromatin structural features from Hi-C matrices, and methods for diagnosing and treating medical conditions or diseases such as cancer. Background Art

[0002] Chromosome conformation capture (often abbreviated as 3C technology or 3C-based methods) is a group of molecular biology methods used to analyze the spatial organization of chromatin in cells. These methods quantify the number of interactions between genomic sites that are adjacent in three-dimensional space but may be separated by many nucleotides in the linear genome. The interaction frequency can be analyzed directly or converted into distance and used to reconstruct the three-dimensional structure.

[0003] High-throughput chromosome conformation capture (Hi-C) allows genome-wide analysis of chromatin interactions in space and has been used to study genome-wide interactions of the genome. It is well known that the spatial organization of chromatin is non-random and is critical for deciphering how the three-dimensional structure of DNA affects genome function and transcription. By comprehensively detecting spatial interactions between genomic regions, Hi-C technology provides deeper insights into the three-dimensional organization of chromatin. Hi-C technology typically involves the generation of hundreds of millions of paired-end sequencing. This technology can capture chromatin interactions throughout the genome and construct a genome-wide Hi-C contact matrix, where each element in the matrix represents the contact strength between any two genomic regions.

[0004] Hi-C uses high-throughput sequencing to find the nucleotide sequence of fragments and uses paired-end sequencing to retrieve short sequences from each end of each ligated fragment. Therefore, for a given ligated fragment, the two sequences obtained should represent two different restriction fragments that were joined together in the proximity-based ligation step. This pair of sequences is aligned to the genome separately to determine the fragments involved in the ligation event. Therefore, all possible pairwise interactions between fragments are tested.

[0005] Currently, Hi-C is the most commonly used method to probe chromatin structure. The paired data generated by this technology are processed through existing general pipelines. The generated data format is a Hi-C matrix, which is a spatial contact frequency matrix between chromatin loci. There are many mathematical methods that can be used to simulate chromatin conformation from Hi-C matrices, including generating spatial distance matrices from Hi-C matrices and / or predicting the coordinates of chromatin loci in three-dimensional space.

[0006] Hi-C is used to calculate multi-scale chromatin structural features to characterize cell states. Existing solutions include: (1) decomposing the Hi-C matrix into a one-dimensional vector through non-negative matrix factorization, which can be used to characterize the topological domains of chromatin; (2) principal component analysis (PCA) of the Hi-C matrix, where the physical meaning of the first principal component (PC1) is the degree to which the DNA bin is in two heterogeneous regions (A / B compartments); (3) starting from the Hi-C matrix, Monte Carlo simulation is performed to obtain the simulated chromatin physical conformation and then extract structural features; (4) using chromatin accessibility data (DNase, MNase) to characterize the structural features of chromatin.

[0007] The disadvantage of existing methods for detecting chromatin transcription start loci, such as chromatin accessibility data, is that they characterize the structural features of chromatin at a relatively small scale (such as the nucleosome scale) and cannot provide structural features at multiple distance scales. The first principal component (PC1), that is, the division of the two arms of chromatin, has the disadvantage of characterizing the structure from the PCA of the Hi-C matrix that the results are unreliable for Hi-C matrices with poor data quality (low sequencing depth and high noise). The disadvantage of performing Monte Carlo simulations from the Hi-C matrix to obtain three-dimensional spatial conformation and structural features is that additional assumptions are required, such as the relationship between contact frequency and physical distance. These assumptions are not necessarily in line with reality, so there are deviations in the calculated conformations and features. The disadvantage of using non-negative matrix decomposition based on the Hi-C matrix to characterize the structure is that it only characterizes the boundaries of topology associated domains (TADs), but not the structural features inside TADs.

[0008] Accurately characterizing the structural features of chromatin loci on a sequence is important for diagnosing and treating diseases with a genetic basis, such as cancer. By looking for specific chromatin interactions that are only present in cancer or only present in normal cells, potential loci associated with cancer can be identified. Therefore, there is a very important need in bioinformatics to find a useful method to identify chromatin structure and structural differences between normal and abnormal cells. These and other problems are addressed by the embodiments disclosed below. Summary of the invention

[0009] The inventors have discovered that chromatin structural features along a DNA sequence can be used to characterize cell types, and that differences in chromatin structure between normal cells and abnormal cells along the sequence can be used to distinguish normal cells from abnormal cells.

[0010] Specifically, the inventors have found that by calculating structural feature vectors along the sequences in the input Hi-C matrix data set and using cell type maps to identify normal cells and abnormal cells, the differences in chromatin structure between different types of cells can be more reliably and effectively discovered. The inventors have also found that such methods can be very useful in diagnosing and treating various medical conditions or diseases, including but not limited to cancer. According to the disclosed embodiments, chromatin structural features can be identified from Hi-C matrices in a novel and unexpectedly effective manner.

[0011] In a first embodiment, a method for identifying chromatin structural features from a Hi-C matrix comprises: performing correlation processing on the Hi-C matrix to calculate a correlation matrix; calculating a structural feature vector based on the correlation matrix; calculating a principal component score matrix based on the structural feature vector; and identifying at least one chromatin structural feature in the principal component score matrix.

[0012] In another embodiment, a non-transitory computer-readable medium stores a program for identifying chromatin structural features from a Hi-C matrix, wherein the program causes a processor to execute: performing correlation processing on the Hi-C matrix to calculate a correlation matrix; calculating a structural feature vector based on the correlation matrix; calculating a principal component score matrix based on the structural feature vector; and identifying at least one chromatin structural feature in the principal component score matrix.

[0013] In another embodiment, a method for diagnosing a medical condition or disease is provided, the method comprising: identifying the chromatin structural feature according to the above method; and correlating the chromatin structural feature with the medical condition or disease.

[0014] In another embodiment, a method for treating a medical condition or disease is provided, the method comprising: identifying the chromatin structural feature according to the above method; and administering a gene therapy vector to a subject in need thereof. The chromatin structural feature indicates a medical condition or disease. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, for ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative labor.

[0016] Figure 1 is a schematic diagram of a method for identifying structural features from a Hi-C matrix according to an embodiment.

[0017] Figure 2 is a schematic diagram of sub-steps of a method for generating an enhanced Hi-C matrix according to an embodiment.

[0018] Figure 3 is a schematic diagram of an algorithm flow for obtaining a structural feature vector from a Hi-C matrix according to an embodiment.

[0019] Figure 4 is a schematic diagram of an algorithm flow of principal component analysis using structural feature vectors according to an embodiment.

[0020] Figure 5 (a) and Figure 5 (b) shows a normal cell according to the embodiment ( Figure 5 (a)) and cancer cells ( Figure 5 (b) Heat map of Hi-C matrix.

[0021] Figure 5 (c) and Figure 5 (d) shows a normal cell according to the embodiment ( Figure 5 (c)) and cancer cells ( Figure 5 (d)) Correlation matrix heat map.

[0022] Figure 5 (e) and Figure 5 (f) shows a normal cell according to the embodiment ( Figure 5 (e)) and cancer cells ( Figure 5 (f) Heat map of the conversion logic matrix.

[0023] Figure 6 A structural feature vector diagram according to an example of an embodiment is shown.

[0024] Figure 7 An exemplary PCA-derived cell type atlas according to an embodiment is shown.

[0025] Figure 8 An exemplary PCA-derived cell type atlas according to an embodiment is shown.

[0026] Fig. 9An exemplary PCA-derived cell type atlas according to an embodiment is shown. DETAILED DESCRIPTION

[0027] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0028] The method of the present invention starts with inputting a Hi-C matrix. A correlation matrix is ​​calculated based on the Hi-C matrix, which represents a description of the similarity between chromatin bins. Based on the correlation matrix, a structural feature vector is calculated. Each element of the structural feature vector is a structural feature of each bin along the DNA sequence. The structural feature vector at multiple scales can be used to characterize the cell state.

[0029] Different cell types have different structural feature vectors. For example, the structural vectors of normal cells and abnormal cells such as cancer cells are clearly distinguishable. The locations of the vector elements with the most significant differences between normal cells and cancer cells are potential targets for cancer detection and anticancer treatment. Specifically, the locations of these elements are the locations of chromatin boxes that undergo significant structural changes during tumorigenesis. Targeting these chromatin boxes can achieve early screening of cancer and discrimination of cancer stages. Chromatin boxes can also be used as targets for anticancer treatment.

[0030] Methods for identifying chromatin structural features from Hi-C matrices

[0031] refer to Figure 1 , the method of the present invention may start with an input Hi-C matrix in step S101. Hi-C matrices are well known in the art. The Hi-C matrix may be raw data, a normalized or enhanced Hi-C matrix. The input may also include a single cell Hi-C matrix or other similar sequencing technology designed to capture chromatin conformation.

[0032] In step S102, a correlation method is applied to the Hi-C matrix to calculate or derive a correlation matrix through multiple iterations, thereby removing noise and strengthening high-order correlations between chromatin bins. The correlation method according to the embodiment is not particularly limited, and may include, but is not limited to, Pearson correlation, Spearman correlation, and cosine similarity. The number of iterations used in the embodiment is also not particularly limited, and may be, for example, in the range of 1 to 20, 1 to 10, or 1 to 4.

[0033] The correlation matrix derived from step S102 can be used as a measure of similarity between chromatin bins for deriving structural feature vectors in step S103. The structural feature vectors can be, for example, one-dimensional structural feature vectors. In an embodiment, the structural feature vectors can be derived by calculating different quantiles of the correlation matrix in step S103a and / or by characterizing the similarity between each site and its sequentially adjacent sites in step S103b.

[0034] In an embodiment, a structural feature vector can be derived by calculating different quantiles of a correlation matrix, as shown in step S103a. In step S103a, different quantiles of a correlation matrix can be calculated and used as a threshold for converting the correlation matrix into a binary matrix. For example, for each quantile, the correlation matrix can be converted into a logical matrix specifically composed of 0 and 1, and the matrix elements greater than the quantile and less than the quantile are converted to 1 and 0, respectively, and vice versa. A one-dimensional vector is obtained by summing the rows of the logical matrix. The one-dimensional vector is a structural feature under a scale limited by a quantile, and its length is the number of chromatin boxes. Structural features under different scales are calculated by different quantiles. The number of quantiles used represents the scale range considered. In other embodiments, step S103a can be implemented by using different specific values ​​instead of different quantiles, and a person of ordinary skill in the art will understand this.

[0035] In an embodiment, as shown in step S103b, the structural feature vector can be derived by characterizing the similarity between each site and its sequentially adjacent sites. In this step, for each site, the average similarity between every 2 sites on the window centered on the site is calculated. For a specified window size and site, a submatrix is ​​extracted from the correlation matrix. Next, for the window size and site, the submatrix is ​​averaged into a one-dimensional vector with a length equal to the number of chromatin boxes. The window size can be selected continuously and can include sites of any suitable range. For example, the window can be in the range of 1 to 10000, 1 to 5000, 1, 1000, 2 to 500, 3 to 100 or 5 to 50 sites. It is understood that different window sizes represent different scales of consideration, and the window size can depend on a variety of factors, including but not limited to the target medical condition or disease and the corresponding genome map.

[0036] The one-dimensional vectors of different scales obtained in step S103 are regarded as structural feature vectors. After calculating the structural feature vector of each sample in step S103, the next step is to calculate the principal component score matrix according to the structural feature vector, as shown in step S104.

[0037] refer to Figure 2, calculating the principal component score matrix based on the structural feature vector may include, in step S104a, for each sample with a feature vector having a defined shape, concatenating the feature vector into an input matrix having a defined shape so that each row of the input matrix is ​​a structural feature vector of the sample, and each column corresponds to a chromatin box along the DNA sequence.

[0038] Calculating the principal component score matrix according to the structural eigenvector may include, in step S104b, normalizing the input matrix, and in step S104c, performing matrix decomposition and dimensionality reduction. Any suitable matrix decomposition and dimensionality reduction method may be performed. In an embodiment, the matrix decomposition and dimensionality reduction method may include but are not limited to PCA, non-negative matrix decomposition and eigenvalue decomposition and singular value decomposition (singular value decomposition, SVD) algorithm. In this step, the rows of the input matrix correspond to the observed values, and the columns correspond to the variables. The initial values ​​of the coefficient matrix and the principal component score matrix are both random matrices. The number of iterative steps can be up to 1000, and can be in the range of 2 to 1000, 5 to 500, 10 to 200 or 20 to 100. The tolerance of the loss function can be up to 1e-6.

[0039] In step S104d, the coefficient matrix and the principal component score matrix are obtained from step S104a. In an embodiment, each column of the coefficient matrix contains the coefficients of a principal component, and the columns are arranged in descending order of component variance. The principal component score matrix is ​​a representation of the input matrix in the principal component space. The rows of the principal component score matrix correspond to samples, and the columns correspond to principal components.

[0040] In step S105, structural features in the principal component scores are identified. In this step, a geometric visualization method can be performed to identify the relationship between the samples from step S104, using the obtained principal component score matrix as a representation of the samples. Any suitable geometric method can be performed. In an embodiment, the geometric method may include, but is not limited to, embedding the principal components into a visualized cell type map.

[0041] The method of the present invention for identifying chromatin structural features from Hi-C matrix will now be described with reference to the following example algorithm to further understand the embodiments of the present invention. However, the present invention is not limited to the specific algorithm described below.

[0042] In an embodiment, the following operations are illustrated using a normalized random non-negative diagonal matrix C, such as Figure 3As shown. For the input matrix C, a Pearson correlation is performed to calculate the correlation matrix of C, namely the correlation matrix D, and four iterations are performed to eliminate noise and strengthen the high-order correlations between chromatin bins. In this case, the correlation matrix D consists of the pairwise linear correlation coefficients between each pair of columns in the input matrix C. The derived correlation matrix D is used as a measure of the similarity between chromatin bins. The correlation matrix D is used as the input to Algorithms I and II described below.

[0043] In Algorithm I, different quantiles of the correlation matrix D are calculated. For each quantile s, the correlation matrix D is converted into a logical matrix I consisting of 0s and 1s. s , converting any matrix elements greater than s and less than s to 1 and 0, respectively. In this case, the 50% quantile s of the correlation matrix D is calculated, which is 0.0132, as shown in Figure 3 As shown. Convert the elements in the correlation matrix D that are greater than the quantile to 1, and the rest to 0. s The row summation of One-dimensional vector is the structural feature at the scale defined by quantile s, and its length is the number of chromatin bins N. Structural features at different scales are calculated according to different quantiles. The number of quantiles used indicates the range of scales considered. The derived vector v is generated t .

[0044] In Algorithm II, the similarity between each site and its sequentially adjacent sites is described. Specifically, for each site i, the average similarity between every two sites in the window centered on i is calculated by the following steps:

[0045] i) For a specific window size w and site i, extract the submatrix from the correlation matrix D Included ab (iw≤a≤i+w,iw≤b≤i+w), where d ab is the element in the a-th row and b-th column of D.

[0046] ii) For a specific window size w and site i, the submatrix Average value and derives a one-dimensional vector of length equal to the number of chromatin bins N

[0047] In this case, the specific window length w is 3, and a submatrix with shape w×w is computed, centered at the i-th element in the diagonal matrix I. The submatrices are averaged to obtain the derived vector v t The i-th element in .

[0048] The one-dimensional vector v at different scales obtained by Algorithms I and IIt is defined as the structural feature vector. Calculate v for each sample t Afterwards, PCA is performed according to Algorithm III.

[0049] In Algorithm III, n samples are sampled. Each sample has a feature vector v of shape 1×N t . All v t Vertically concatenate into a matrix X with shape n×N, so that each row of input X is the structural feature vector v of the sample t , where each column corresponds to a chromatin bin along the DNA sequence. Then, proceed as follows:

[0050] i) Normalize X.

[0051] ii) Perform PCA on X. The rows of X correspond to observations and the columns correspond to variables. The SVD algorithm is used. The initial values ​​of the coefficient matrix coeff and the principal component score matrix score are random matrices. The maximum number of iterations allowed is 1000. The tolerance of the loss function is 1e-6.

[0052] iii) Obtain the coefficient matrix coeff and score. The coefficient matrix coeff is N×N, and the principal component score matrix score is n×p, where p is the number of principal components. Each column of coeff contains the coefficient of a principal component, and the columns are arranged in descending order of component variance. The principal component score matrix score is a representation of X in the principal component space. The rows of score correspond to samples, while the columns correspond to principal components.

[0053] Then, geometric methods are used to find the relationship between the samples to obtain the score as a representation of the sample. Figure 4 The algorithm flow of principal component analysis using structural eigenvectors is shown. Figure 4 In the example, each row of the input matrix corresponds to the structural feature vector obtained above. The structural feature vectors of 11 samples are calculated, each of which has a length of 4, and the input matrix X is 11×4, that is, the number of chromatin boxes N=4. It can be understood that the number of chromatin boxes N is not particularly limited, and can be, for example, in the range of 100 to 20,000, 500 to 10,000, or 1,000 to 6,000.

[0054] Structural feature vector v t Each element of is a structural feature of each box along the DNA sequence. In this case, cancer cells refer to various solid tumor cancers, leukemias, and various cancer cell lines and leukemia cell lines, while normal cells refer to the corresponding normal cells of a specific cancer. Figure 4 The structural feature vector v t The clustering is very obvious.

[0055] Methods for diagnosing and treating medical conditions or diseases

[0056] In other embodiments, methods for diagnosing and treating medical conditions or diseases are provided. These methods include identifying chromatin structural features in the above-mentioned visualized cell type atlas. The visualized cell type atlas obtained also allows to distinguish normal cells from cancer cells, describe the development of cancer, and distinguish different cancers, which is useful for targeting and treating medical conditions or diseases such as cancer. Treatment may include using the transcription or translation product of the obtained site as a medical condition or disease target. This step may include identifying at least one site associated with structural chromatin aberrations in the target cell. At least one site may include, but is not limited to, SPAG9, TOB1, and UTP18.

[0057] In a method for diagnosing a disease, the chromatin structural feature is indicative of the disease. In a method for treating a disease, the method comprises administering a gene therapy vector to a subject in need thereof. The gene therapy may comprise using a transcription or translation product of at least one site associated with the chromatin structural feature in a target cell as a disease target.

[0058] According to the method of the present invention, regulatory genes or regulatory elements that can regulate the sequence of the open sequencing frame can be identified by physical interaction (close spatial proximity) between these regulatory elements and these open sequencing frames. Regulatory elements and open sequencing frames can be close or far apart along the linear genomic sequence, and can also be located on different chromosomes. The open sequencing frame sequence can be associated with a medical condition or disease.

[0059] In particular, sites that are prone to change in medical conditions or diseases such as cancer can be found as targets for disease diagnosis and treatment. The inventors found that different types of cancer samples showed highly consistent features, indicating that the method is surprisingly effective in identifying common features of cancer cell structures and provides new ideas for cancer diagnosis and treatment.

[0060] The disclosed embodiments are applicable to and can operate on any medical condition or disease that has a genetic basis. In this regard, medical conditions or diseases may include, but are not limited to, cancer, cardiovascular disease, kidney disease, autoimmune disease, lung disease, liver disease, lymphatic disease, bone marrow disease, bone disease, blood disease, etc.

[0061] Non-transitory computer-readable media and machine learning

[0062] Embodiments of the present invention also include non-transitory computer readable media storing a program for identifying chromatin structural features from a Hi-C matrix, which causes a processor to perform the method of the present invention. Embodiments of the present invention may also include various machine learning algorithms implemented on a dedicated computer or computer system for performing any one or more of the methods of the present invention. In this regard, the algorithm can be used to automatically perform steps using commercial or open source tools. Machine learning algorithms can be used to mathematically process large genomic data sets and can also be used to optimize calculations and improve the precision and accuracy of outputs.

[0063] As understood in the field of bioinformatics, machine learning algorithms involve building classifiers and training datasets. Classifiers play an important role in analyzing complex multidimensional systems such as chromatin structure and eukaryotic genomes. To develop classification, supervised learning techniques can be based on decision trees, logical rules, or other mathematical techniques such as linear discriminant methods (including perceptrons, support vector machines, and related variants), nearest neighbor methods, Bayesian inference, neural networks, etc.

[0064] The programming tools used to develop the disclosed machine learning algorithms are not particularly limited and may include, but are not limited to, open source tools such as The rules engine includes Programming languages ​​such as SQL, R, Matlab and Python, and various relational database architectures. In an embodiment, Python is the preferred programming framework in which the method of the present invention is performed.

[0065] The special-purpose computer or processing system on which the methods and machine learning algorithms of the present invention may be implemented may be a special-purpose processing system and may operate with many other general-purpose or special-purpose computing system environments or configurations, as understood by bioinformatics practitioners. Examples of known computing systems, environments, and / or configurations that may be suitable for use with the methods of the present invention may include, but are not limited to, personal computer systems, server computer systems, thin clients, thick clients, handheld or laptop devices, multiprocessor systems, microprocessor-based systems, set-top boxes, programmable consumer electronics, network PCs, minicomputer systems, mainframe computer systems, and distributed cloud computing environments including any of the above systems or devices, etc.

[0066] Computer systems may be described in the general context of computer system executable instructions, such as program modules, executed by the computer system. Generally, program modules may include routines, programs, objects, components, logic, data structures, etc. that perform specific tasks or implement specific abstract data types. Computer systems may be implemented in a distributed cloud computing environment, where tasks are performed by remote processing devices linked through a communications network. In a distributed cloud computing environment, program modules may be located in local and remote computer system storage media, including memory storage devices.

[0067] A neural network may be used to perform the method of the present invention. The neural network may be a deep convolutional neural network. The neural network may be a deep neural network comprising an output layer and one or more hidden layers. In an embodiment, training the neural network may include training the output layer by minimizing a loss function for a given optimal allocation set, and training the hidden layer by a back propagation algorithm.

[0068] The deep neural network can be a convolutional neural network (CNN). In a CNN-based model, a set of filters is used to extract features through convolution operations. The training of the CNN is done using a training data set, which determines the training values ​​of the neural network parameters / weights.

[0069] In some CNN models, the number of CNN layers and fully connected layers can vary. In some network architectures, residual transfer or feedback can be used to avoid the conventional problem of gradient disappearance when training network weights. The network can be built using any suitable computer language, such as Python or C++. Deep learning toolboxes, such as TensorFlow, Caffe, Keras, Torch, Theano, CoreML, etc., can be used to implement the network. These toolboxes are used to train the weights and parameters of the network. In some embodiments, customized implementations of CNN and deep learning algorithms on dedicated computers with graphics processing units (GPUs) are used for training, reasoning, or both. The reasoning refers to the stage of using the trained model to infer / predict the test sample. The weights of the trained model are stored on the computer disk and then used for reasoning. Different optimizers, such as the Adam optimization algorithm and different optimizers for gradient descent, can be used to train the weights and parameters of the network. When training the network, hyperparameters can be adjusted to achieve higher recognition and detection accuracy. During the training phase, the network can be exposed to training data through several epochs. The artificial intelligence training pattern is defined as the entire data set being passed forward and backward only once through the neural network.

[0070] The network can be trained using a transfer learning mechanism. In transfer learning, the weights of the network are first trained using a dataset different from the target dataset to learn relevant features. This pre-trained network is then further retrained using features from the target database. The CNN architecture can be three-dimensional to process three-dimensional chromatin structure data.

[0071] Example

[0072] Example 1

[0073] Calculate the correlation matrix from the Hi-C matrix to distinguish three normal oral samples from three cancer oral samples, as Figure 5 (a)-5(f). First, generate a Figure 5 (a)) and cancer cells ( Figure 5 (b) Heat map of the original unprocessed Hi-C matrix. Figure 5 (a) and 5(b), the distinction between normal and tumor samples is unclear.

[0074] Figure 5 (c) and 5(d) correspond to Figure 5 (a) and 5(b), show the heatmaps after applying the Pearson correlation matrix of the fourth iteration. Figure 5 (c) and 5(d), the differences between normal samples and cancer samples are significant in the calculated correlation matrices.

[0075] Figure 5 (e) and 5(f) correspond to Figure 5 (c) and 5(d) show the heatmaps of the transformed logical matrix (consisting of 0s and 1s) obtained by using the 50% quantile of the previous correlation matrix as the threshold. Figure 5 (e) and 5(f), the distinction between normal and cancer samples is significant in the calculated logistic matrices.

[0076] Example 2

[0077] Generate a one-dimensional structural feature vector v for each of the three normal oral samples (i.e., N1, N2, and N3) and three cancer oral samples (i.e., T1, T2, and T3) on chromosome 22 t The results are shown in Fig. Figure 6 As shown. Figure 6 As shown, cancer samples are easily distinguished from normal samples.

[0078] Example 3

[0079] For the one-dimensional structural feature vector v of 19 samples t PCA was performed to generate a PCA-derived cell type map. Figure 7 As shown in . The first two dimensions of the principal components obtained are used to represent the samples. Figure 7 In the figure, each line represents the v of a sample t The lines composed of + and o represent the v of cancer and normal samples, respectively. t The structural feature vectors v of 19 samples involving three types of cancer (oral cancer, colon cancer, and bladder cancer) are shown. t Value. Figure 7 The resulting cell type maps clearly distinguished normal and cancer samples, as well as samples of different cancers.

[0080] Example 4

[0081] For the one-dimensional structural feature vector v of 33 samples t PCA was performed to generate a PCA-derived cell type map. Figure 8 As shown. The first two dimensions of the principal components obtained are used to represent the samples. Among the 33 samples, there are 4 normal blood samples (indicated by "o"), 22 leukemia samples (indicated by "*"), and 7 leukemia cell line samples (indicated by "+"). Figure 8 As shown, normal blood, leukemia, and leukemia cell line samples are clearly shown as three clusters, and each cell type is distinguishable from the other types.

[0082] Example 5

[0083] The one-dimensional structural feature vector v of 36 cancer samples t PCA was performed to generate a PCA-derived cell type map. Fig. 9 As shown. The first two dimensions of the principal components obtained are used to represent the samples. Among the 36 cancer samples, there are 21 leukemia samples (represented by diamonds), 7 oral cancer samples (represented by "*"), 3 colon cancer samples (represented by "o"), 3 bladder cancer samples (represented by five-pointed stars) and 2 lung cancer samples (represented by "+"). Fig. 9 As shown, the five cancer types are clearly shown as five clusters, and each cell type is distinguishable from the other types.

[0084] It should be understood that the above disclosed features and functions or their alternatives may be desirably combined into different devices, systems and methods. In addition, various substitutions, modifications, changes or improvements may be made subsequently by those skilled in the art, and are also intended to be covered by embodiments of the present invention. Therefore, various changes may be made without departing from the spirit and scope of the present invention.

Claims

1. A method for identifying chromatin structural features from a Hi-C matrix, the method comprising: Performing correlation processing on the Hi-C matrix to calculate a correlation matrix; Based on the correlation matrix, calculating a one-dimensional structural feature vector; The calculating of the one-dimensional structural feature vector based on the correlation matrix includes calculating the quantile of the correlation matrix, and calculating the average similarity between adjacent sites in a window for each site, generating a submatrix according to the correlation matrix based on the size of the window, and averaging the submatrix into the one-dimensional structural feature vector having a length equal to the number of chromatin boxes; wherein the calculating of the quantile of the correlation matrix includes converting the correlation matrix into a binary matrix, converting matrix elements greater than the quantile into 1 or 0, and converting matrix elements less than the quantile into the other of 1 or 0; According to the one-dimensional structural feature vector, a principal component score matrix is ​​calculated; the calculation of the principal component score matrix according to the one-dimensional structural feature vector includes: splicing the one-dimensional structural feature vector into an input matrix with a limited shape so that each row of the input matrix is ​​a one-dimensional structural feature vector; normalizing the input matrix; performing matrix decomposition and dimensionality reduction to obtain a coefficient matrix and the principal component score matrix; Identify at least one chromatin structural feature in the principal component score matrix; the identifying at least one chromatin structural feature in the principal component score matrix includes performing geometric visualization, and the geometric visualization is a visualized cell type map.

2. The method for identifying chromatin structural features from a Hi-C matrix according to claim 1, wherein: The Hi-C matrix is ​​the original data Hi-C matrix.

3. The method for identifying chromatin structural features from a Hi-C matrix according to claim 1, wherein: The Hi-C matrix is ​​a normalized Hi-C matrix.

4. The method for identifying chromatin structural features from a Hi-C matrix according to claim 1, wherein: The correlation processing is at least one correlation processing selected from Pearson correlation, Spearman correlation and cosine similarity.

5. The method for identifying chromatin structural features from a Hi-C matrix according to claim 1, wherein: Performing matrix decomposition and dimensionality reduction includes selecting at least one of principal component analysis, non-negative matrix decomposition, eigenvalue decomposition, and singular value decomposition algorithms.

6. A non-transitory computer readable medium storing a program for identifying chromatin structural features from a Hi-C matrix, the program causing a processor to execute: Performing correlation processing on the Hi-C matrix to calculate a correlation matrix; Based on the correlation matrix, a one-dimensional structural feature vector is calculated; the calculation of the one-dimensional structural feature vector based on the correlation matrix includes calculating the quantile of the correlation matrix, and calculating the average similarity between adjacent sites in a window for each site, generating a submatrix according to the correlation matrix based on the size of the window, and averaging the submatrix into the one-dimensional structural feature vector with a length equal to the number of chromatin boxes; wherein the calculation of the quantile of the correlation matrix includes converting the correlation matrix into a binary matrix, converting matrix elements greater than the quantile into 1 or 0, and converting matrix elements less than the quantile into the other of 1 or 0; According to the one-dimensional structural feature vector, a principal component score matrix is ​​calculated; the calculation of the principal component score matrix according to the one-dimensional structural feature vector includes: splicing the one-dimensional structural feature vector into an input matrix with a limited shape so that each row of the input matrix is ​​a one-dimensional structural feature vector; normalizing the input matrix; performing matrix decomposition and dimensionality reduction to obtain a coefficient matrix and the principal component score matrix; Identify at least one chromatin structural feature in the principal component score matrix; the identifying at least one chromatin structural feature in the principal component score matrix includes performing geometric visualization, and the geometric visualization is a visualized cell type map.