Inter-cell communication inference method based on single cell transcriptome data
By applying a joint learning model of matrix decomposition and sparse constraints in single-cell transcriptome data, the problem that existing tools cannot predict intra- and external communication pathways is solved, and an accurate inference of intercellular communication networks and in-depth understanding of cellular functions is achieved.
Patent Information
- Application Number
- CN202411032469.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-07-30
- Publication Date
- 2024-11-08
AI Technical Summary
Existing intercellular communication network inference tools lack the ability to predict communication pathways that connect inside and outside the cell, and cannot fully understand the complexity and dynamic nature of intercellular communication.
Using a method based on single-cell transcriptome data, a joint learning model is constructed through bioinformatics, computational biology and machine learning algorithms, combining matrix decomposition and sparse constraints, gene functional module selection and cell clustering are carried out, and intercellular and cell integration are integrated. communication signals within the cell communication network.
Accurate inference of intercellular communication networks is achieved, gene expression patterns and cell dynamic processes are revealed, inferring performance of intercellular communication networks is improved, and a deep understanding of cell functions and regulatory mechanisms are provided.
Smart Images

Figure CN118918962A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of bioinformatics, and in particular to a method for inferring intercellular communication based on single-cell transcriptome data. Background Art
[0002] Cells are the basic units that make up organisms, and the development of single-cell transcriptome sequencing technology has promoted a deeper understanding of the diversity of cell states. In multicellular organisms, intercellular communication enables multiple cells to work in coordination to form tissues, organs or systems to complete various biological tasks. Intercellular communication can more accurately analyze the functional differences, mutual recognition, and signal transmission mechanisms between different cell types, and thus gain a deeper understanding of the complexity of the internal states and relationships of cells. Through the study of cell communication, the details of the microscopic interactions between cells are revealed, which helps to identify how specific cell populations respond under different physiological and pathological conditions, as well as their roles in the development of diseases. This is of great significance for personalized medicine, disease diagnosis, and the formulation of treatment strategies.
[0003] Although various computational tools have been developed to infer intercellular communication, these tools mainly focus on the expression intensity and specificity of ligand-receptor pairs, and infer intercellular communication by calculating the co-expression, differential expression and expression correlation of ligands and receptors. However, there are certain limitations in inferring intercellular communication based on the expression intensity of ligand and receptor pairs. First, some receptors are usually expressed at a relatively stable level in cells and are not directly involved in the regulation of intercellular communication, so they are not closely related to intercellular communication. Second, the mRNA encoding some surface receptors is usually low in abundance, which may lead to the inability to detect these receptors at the single cell level. In addition, intercellular communication includes not only signal transduction between intercellular ligands and receptors, but also the transmission and amplification of intracellular signals through specific signal pathways, which ultimately leads to changes in the activity of downstream transcription factors and gene regulatory networks. Therefore, a comprehensive understanding of the relevant communication signals inside cells and their highly integrated and dynamic nature can accurately reveal the interactions between different cell types and the signal transduction network between cells, help researchers discover more hidden connections and interactions between cells, and deepen their understanding of cell functions and regulatory mechanisms.
[0004] The complexity and heterogeneity of single-cell data limit the comprehensive characterization of gene expression patterns and cellular dynamic processes. Extracting key features from single-cell data, identifying functional modules, and clustering cells according to their similar features are crucial for a comprehensive understanding and interpretation of the mechanisms and patterns of intercellular communication. A joint learning model based on matrix decomposition is used to select features and detect functional gene modules to explore gene expression patterns, while clustering is performed to discover cell identity and function. By comprehensively mining the information in single-cell data, the consistency and accuracy of data analysis are guaranteed, and the problem of result interpretation caused by differences between different tools is avoided, so that intercellular communication can be analyzed and interpreted more accurately. Combined with the annotation information of cell identity, the relevant communication signals between cells and within cells are integrated to achieve the inference of the communication network between cells. By accurately inferring the modes and mechanisms of intercellular communication, the patterns and associations hidden in the data can be better revealed, thereby deeply understanding the interactions between cells. This is of great significance for understanding the process of disease occurrence and development, discovering new treatments, and designing drug targets based on cell signaling. Summary of the invention
[0005] The purpose of the present invention is to solve the problem that existing intercellular communication network inference tools lack the ability to predict the communication pathways connecting the inside and outside of cells, and to propose an intercellular communication inference method based on single-cell transcriptome data. Through the combination of bioinformatics, computational biology and machine learning algorithms, accurate inference of the intercellular communication network of single-cell transcriptome data can be achieved.
[0006] In order to achieve the above object, the technical solution adopted by the present invention is:
[0007] A method for inferring intercellular communication based on single-cell transcriptome data, comprising the following steps:
[0008] In the single-cell data, the genes with low expression levels are filtered, i.e., the genes expressed in less than a set percentage of cells, and then the genes with significant expression differences between different cells are selected, i.e., the genes with the highest standardized variance between different cells, i.e., the first set number of genes, and finally, regularization preprocessing is performed, and these single-cell data are recorded in the form of a matrix;
[0009] The single-cell data matrix after regularization preprocessing is decomposed into a gene projection matrix, a cell projection matrix and an association matrix, which respectively represent the expression of genes and cells in the projection space and the association between genes and cells. The cell projection matrix is further processed using a non-negative matrix decomposition method to obtain a coefficient matrix, which represents the cell expression in a low-dimensional space. Combining the two decomposition methods, and applying sparse constraints and graph regularization constraints to the cell projection matrix and the coefficient matrix, a joint learning model based on matrix decomposition is proposed.
[0010] A joint learning model based on matrix decomposition is used to achieve joint learning of multiple single-cell analysis tasks: the gene projection matrix is used to select representative genes and determine gene function modules; the association matrix is used to calculate the importance of gene function modules and sort the gene function modules; the coefficient matrix is used to cluster single-cell data and compare them with known cell marker genes to identify the cell type represented by each cluster;
[0011] Based on the results of cell type identification, the scores of communication signals transmitted by ligands and receptors between different types of cells and the scores of communication signals transmitted by downstream transcription factors and their target genes within cells are calculated respectively; by integrating intercellular and intracellular communication signals, an intercellular communication inference model based on single-cell transcriptome data is obtained.
[0012] Preferably, in the single-cell data, genes with low expression levels are filtered, and then genes with significant expression differences between different cells are selected, and finally regularization preprocessing is performed, including the following steps:
[0013] 1.1) First, the genes with low expression in the single-cell data are filtered. These genes are expressed in less than 6% of the cells (6% by default) and are considered to be unable to provide useful information in the clustering process. Such genes are filtered;
[0014] 1.2) Secondly, for the filtered single-cell data, the principal component analysis method was used to select the top 2000 genes with the highest standardized variance between different cells, and log 2 Regularization for downstream analysis:
[0015] X i,j =log 2 (M i,j +1) (1)
[0016] Among them, M i,j represents the expression level of the i-th gene in the j-th cell in the original single-cell data, X i,j Represents the expression level of the i-th gene in the j-th cell in the normalized single-cell data.
[0017] Preferably, the single-cell data matrix after regularization preprocessing is decomposed by projection matrix to obtain gene projection matrix, cell projection matrix and association matrix, which respectively represent the expression of genes and cells in the projection space and the association between genes and cells; the cell projection matrix is further processed by non-negative matrix decomposition method to obtain coefficient matrix, which represents the cell expression in low-dimensional space; the two decomposition methods are combined, and sparse constraints and graph regularization constraints are respectively applied to the cell projection matrix and the coefficient matrix, and a joint learning model based on matrix decomposition is proposed, including the following steps:
[0018] 2.1) Use projection matrix decomposition method for feature selection, map the original data into a low-dimensional representation space, which can effectively reduce the dimension of the data while retaining key gene feature information; decompose the input single-cell gene expression matrix X into gene projection matrix U, cell projection matrix V and correlation matrix S, so as to calculate the difference between the original features and the selected features:
[0019]
[0020] in, represents the Froben i us norm; X∈R m×n Represents the input single-cell gene expression matrix, containing m genes and n cells; gene projection matrix and the cell projection matrix Respectively represent the expression of genes and cells in the projection space, m is the number of genes, and n is the number of cells; the association matrix It provides additional degrees of freedom to ensure the accuracy of the projection matrix, and k1 is the dimension after decomposition;
[0021] 2.2) In the process of feature selection, for the cell projection matrix V, a sparse constraint term L is introduced 2,1 norm to eliminate the inherent sparsity of single-cell data and thus improve the accuracy of the algorithm; L 2,1 The norm is first calculated as the L of the row vector 1 norm, which helps to eliminate the sparsity of single-cell data and thus improve the interpretability and accuracy of the algorithm. Subsequently, L is applied to the column vector 2 The calculation of the norm can effectively reduce the impact of noise and outliers on the algorithm, thereby improving the robustness of the algorithm:
[0022]
[0023] Among them, the data element v ij Represents the element value of the i-th row and j-th column in the matrix V; the target formula of feature selection is rewritten as:
[0024]
[0025] 2.3) Cluster the cells using non-negative matrix factorization method; the non-negative matrix factorization method decomposes the matrix V into a base matrix B and a coefficient matrix F:
[0026]
[0027] in, They are represented as basis matrix and coefficient matrix, respectively, m is the number of genes, n is the number of cells, k 2 is the dimension of dimensionality reduction;
[0028] 2.4) Using graph regularization methods to constrain the non-negative matrix factorization process helps to better capture the potential relationship between data, making the clustering results more robust and reliable: Construct a graph G to describe the proximity between cells in the coefficient matrix F, where the nodes on the graph represent cells and the edges on the graph represent the similarities between cell pairs; define a symmetric weight matrix W to quantify the edges, where the elements in W represent the weights of the edges connecting cell i and cell j:
[0029]
[0030] Among them, v i is the ith cell, N K (v i ) is v i The set of K nearest neighbors of ; then the formula of the graph regularization constraint is as follows:
[0031]
[0032] The objective formula of non-negative matrix factorization is further rewritten as:
[0033]
[0034] Where D is a diagonal matrix, and the elements in D are obtained by summing the rows or columns of W; f i and f j Yes i and v j The low-dimensional representation of G; L is the Laplacian matrix of G, that is, L = DW;
[0035] In order to solve the independence problem of feature selection and cell clustering, projection matrix decomposition and non-negative matrix decomposition are combined to jointly learn feature selection and cell clustering. Combining formula (4) for feature selection and formula (8) for clustering, a joint learning model based on matrix decomposition method is constructed, and its objective function is defined as the following optimization problem:
[0036]
[0037] The objective formula is solved using an iterative strategy to optimize one variable by fixing the other variables until the termination criterion is reached.
[0038] Preferably, a joint learning model based on matrix decomposition is used to realize joint learning of multiple single-cell analysis tasks: representative genes are selected and gene function modules are determined using a gene projection matrix; the importance of gene function modules is calculated using an association matrix and the gene function modules are sorted; single-cell data are clustered using a coefficient matrix and compared with known cell marker genes to identify the cell type represented by each cluster, including the following steps:
[0039] 3.1) Using the cell projection matrix U obtained by the joint learning model based on matrix decomposition, the features with expression values greater than the specified threshold in each column are selected as members of the module, thereby constructing a gene function module; for each column vector u i ,i=1,...,k, select the nodes whose weighted factor absolute value is greater than the threshold θ according to their z-score:
[0040]
[0041] in, is the mean of the row vectors of the matrix U, is the standard deviation of the row vector of U; based on the above transformation, the condition z will be satisfied ij >θ's g(i) is regarded as a member of the i-th module, and the threshold θ is set to the value corresponding to the highest point on the module similarity curve;
[0042] 3.2) The importance of the module is quantified by the association matrix S obtained by the joint learning model based on matrix decomposition; the element s of the association matrix S is ij As weights, u is used to reconstruct the matrix X i and v j A linear combination of ij The larger the value, the i and v j The higher the similarity between the identified modules, the greater the element values of all combinations of the selected features in the matrix X; the correlation matrix S is a diagonal matrix whose non-zero elements are only on the diagonal. The importance of the identified modules is determined by evaluating the value of the matrix S;
[0043] 3.3) Generate the coefficient matrix F using a joint learning model based on matrix decomposition, and assign cells to cell clusters based on the maximum coefficient of the matrix F; at the same time, combine the expression levels of marker genes in specific cell types to label the cell clusters as known cell types; this method can effectively identify and interpret cell clusters, and help to more fully understand the identity and function of cells. For cell clusters that lack known marker gene expression information, assign them to the known cell type with the largest number of cells;
[0044] Preferably, based on the results of cell type identification, the scores of the communication signals transmitted between different types of cells by ligands and receptors and the scores of the communication signals transmitted within cells by downstream transcription factors and their target genes are calculated respectively; by integrating the intercellular and intracellular communication signals, an intercellular communication inference model based on single-cell transcriptome data is obtained, comprising the following steps:
[0045] 4.1) Collect and integrate three layers of data information to build a reference library of intercellular communication DcjCommDB as evidence for inferring cell-to-cell interactions, including: ligand-receptor (LR) interactions, transcription factor-target gene (TF-TG) interactions, and signal transduction from receptors to transcription factors;
[0046] 4.2) Intercellular signaling fraction through ligand-receptor (LR) interaction 2 norm to evaluate, that is LR k is a two-dimensional vector represented by the normalized expression values of the ligand and receptor of a ligand-receptor (LR) interaction:
[0047]
[0048] Among them, L i,k is the average expression value of ligand L in cell type i, R j,k is the average expression value of receptor R in cell type j; if ligand L contains subunits and represents the expression value of subunits, it is defined as the geometric mean of the expression values of all subunits: Likewise, the receptor R is defined as l g and r h are the expression values of g subunit and h subunit in the ligand and receptor, respectively;
[0049] 4.3) Combine the interactions between receptors, transcription factors and target genes to calculate the intracellular communication signal score S intra :
[0050]
[0051] Among them, TF PPR(t,r,p) represents the interaction between receptor r and transcription factor t in pathway p; TF activity (t,k) represents the activity of each transcription factor t in cell type k calculated according to Fisher's test:
[0052] TF activity (t,k)=1-P Fisher (t,k) (13)
[0053] 4.4) Combined intercellular communication fraction S inter and intracellular communication fraction S intra , calculate the cell interactions S between different cell types k , which is defined as:
[0054] S k =S inter ×S intra (14).
[0055] Among them, the DcjCommDB reference library collects 5597 human ligand-receptor (LR) interactions, 18421 TGs regulated by 1537 TFs, and 411 TFs regulated by 1533 receptors. It also collects 5818 mouse ligand-receptor (LR) interactions, 1457 TGs regulated by 1649 TFs, and 369 TFs regulated by 731 receptors.
[0056] Beneficial effects of the present invention:
[0057] The present invention utilizes the communication signals between ligand-receptor pairs, transcription factors and their target genes to achieve communication between the inside and outside of cells, thereby accurately inferring the communication network between cells. By utilizing a joint learning model based on matrix decomposition and an inference statistical model of intercellular communication, the complexity and heterogeneity of single-cell data are solved, thereby effectively exploring the expression pattern of genes, and discovering the identity and function of cells, further enhancing the reasoning performance of the intercellular communication network. The beneficial effects of the present invention specifically include the following:
[0058] 1. The present invention performs multiple single-cell data analysis tasks in a coherent manner, comprehensively revealing gene expression patterns and cell dynamic processes. By analyzing the gene expression data of individual cells, gene sets with similar expression patterns are identified, thereby revealing complex functional modules and signal pathways within cells. By selecting representative features and performing cell clustering analysis, the similarities and differences between cells are revealed, thereby inferring the communication and interaction relationships between cells.
[0059] Second, the present invention reveals the regulatory mechanism and potential key factors of intracellular biological processes by identifying gene function modules with similar expression patterns. Using the gene projection matrix obtained based on the matrix decomposition joint learning model, gene sets with similar expression patterns can be identified, thereby revealing the complex gene function modules in cells, providing important clues for a deeper understanding of the internal regulatory mechanisms of various biological processes in cells.
[0060] 3. The present invention utilizes the features and similarity relationships in the data, selects representative features and performs cell clustering, and reveals the cell type, state and structural information hidden in the data. By combining sparse constraints and graph regularization constraints, the joint learning model based on matrix decomposition maintains the consistency of the algorithm while reducing the impact of noise. By formulating the joint learning of feature selection and cell clustering as a constrained optimization problem and deriving the corresponding optimization rules, the accuracy of cell type discovery is significantly improved, and a deeper understanding of cell diversity and functional characteristics is promoted.
[0061] Fourth, the present invention comprehensively and accurately infers the communication signals between cells and within related cells by revealing the role of ligand-receptor pairs, transcription factors and their target genes in cell communication. By predicting and visualizing intercellular communication and analyzing how these signals coordinate functions, it is helpful to gain a deeper understanding of the regulatory process of life activities and the mechanism of disease occurrence. The scheme of the present invention is independently innovative. By comprehensively analyzing single-cell data, it helps to have a deeper understanding of the composition and function of cell populations, providing a more comprehensive and in-depth perspective for cell biology and disease research. BRIEF DESCRIPTION OF THE DRAWINGS
[0062] Figure 1 It is a flow chart of the intercellular communication inference method DcjComm based on single-cell transcriptome data of the present invention;
[0063] Figure 2 It is a flow chart of constructing the intercellular communication network inference method DcjComm of the present invention;
[0064] Figure 3 This is a feature selection result diagram of the intercellular communication network inference method DcjComm of the present invention;
[0065] Figure 4 It is a performance result diagram of the gene function module detected by the intercellular communication network inference method DcjComm of the present invention;
[0066] Figure 5 It is a cell clustering result diagram of the intercellular communication network inference method DcjComm of the present invention;
[0067] Figure 6This is a result diagram of the intercellular communication network predicted by the intercellular communication network inference method DcjComm of the present invention on single-cell transcriptome data. DETAILED DESCRIPTION
[0068] A specific implementation of the present invention is described in detail below in conjunction with the accompanying drawings, but it should be understood that the protection scope of the present invention is not limited by the specific implementation.
[0069] See also Figure 1-Figure 6 , the present invention discloses a method for inferring intercellular communication based on single-cell transcriptome data.
[0070] See also Figure 1 As shown, the steps of the method of the present invention are mainly summarized as follows: 1. Single-cell data preprocessing, which includes: ① Select highly variable genes, ② Regularize; 2. Joint learning model based on matrix decomposition, which includes: ① Construct a joint learning model, ② Optimize the model; 3. Identify gene expression patterns and cell subpopulations, which includes: ① Select representative features, ② Detect gene function modules, ③ Perform cell clustering; 4. Construct an inference statistical model of intercellular communication, which includes: ① Construct an intercellular communication network, ② Evaluate model performance. The present invention overcomes the problem that the communication pathways connecting the inside and outside of cells are difficult to predict. Through an in-depth understanding of the internal states and relationships of cells, the performance of the inference of the intercellular communication network is improved, which helps to more fully understand the underlying biological processes. The method of the present invention is detailed as follows:
[0071] The method for inferring intercellular communication based on single-cell transcriptome data provided by the present invention specifically comprises the following steps:
[0072] S1: Single-cell data preprocessing: This step aims to preprocess the single-cell transcriptome data, including filtering out genes with abnormally low expression levels, selecting genes with significant expression differences between different cells, and performing regularization preprocessing. Through these steps, a more reliable data foundation is provided for subsequent model construction and analysis;
[0073] S1.1 First, we filter genes with too low expression in the single-cell data. These genes are expressed in less than α% of cells (α=6 by default) and are generally considered to provide no useful information in the clustering process.
[0074] S1.2 Secondly, for the filtered single-cell data, the principal component analysis method was used to select the top 2000 genes with the highest standardized variance between different cells, and log 2 Regularization for downstream analysis.
[0075] X i,j =log 2 (Mi,j +1) (1)
[0076] Among them, M i,j represents the expression level of the i-th gene in the j-th cell in the original single-cell data, X i,j Represents the expression level of the i-th gene in the j-th cell in the normalized single-cell data.
[0077] S2: Construction of joint learning model based on matrix decomposition: Figure 2 As shown in the figure, the preprocessed single-cell data matrix is decomposed by projection matrix to obtain gene projection matrix, cell projection matrix and association matrix, which respectively represent the expression of genes and cells in the projection space and the association between genes and cells; the cell projection matrix is further processed by non-negative matrix decomposition method to obtain coefficient matrix, which represents the cell expression in low-dimensional space; combining the two decomposition methods, and applying sparse constraints and graph regularization constraints to the cell projection matrix and coefficient matrix respectively, a joint learning model based on matrix decomposition is proposed;
[0078] S2.1 uses the projection matrix decomposition method for feature selection, mapping the original data into a low-dimensional representation space, effectively reducing the dimension of the data while retaining key gene feature information. The input single-cell gene expression matrix X is decomposed into the gene projection matrix U, the cell projection matrix V and the correlation matrix S, thereby calculating the difference between the original features and the selected features:
[0079]
[0080] in, represents the Froben i us norm. X∈R m×n Represents the input single-cell gene expression matrix, which contains m genes and n cells. Gene projection matrix and the cell projection matrix Respectively represent the expression of genes and cells in the projection space, m is the number of genes, and n is the number of cells. Additional degrees of freedom are provided to ensure the accuracy of the projection matrix, and k1 is the dimension after decomposition.
[0081] S2.2 In the process of feature selection, for the cell projection matrix V, a sparse constraint term L is introduced 2,1 norm to eliminate the inherent sparsity of single-cell data and thus improve the accuracy of the algorithm. 2,1 The norm is first calculated as the L of the row vector 1 norm, which helps to eliminate the sparsity of single-cell data and thus improve the interpretability and accuracy of the algorithm. Subsequently, L is applied to the column vector 2The calculation of the norm can effectively reduce the impact of noise and outliers on the algorithm, thereby improving the robustness of the algorithm:
[0082]
[0083] Among them, the data element v ij Represents the element value of the i-th row and j-th column in the matrix V. The target formula of feature selection can be rewritten as:
[0084]
[0085] S2.3 uses non-negative matrix factorization to cluster cells. The non-negative matrix factorization method decomposes the matrix V into a basis matrix B and a coefficient matrix F:
[0086]
[0087] in, They are represented as basis matrix and coefficient matrix, respectively, m is the number of genes, n is the number of cells, k 2 is the dimension of dimensionality reduction.
[0088] S2.4 uses graph regularization methods to constrain the non-negative matrix factorization process, which helps to better capture the potential relationships between data, making the clustering results more robust and reliable. Construct a graph G to describe the proximity between cells in the coefficient matrix F. The nodes on the graph represent cells, and the edges on the graph represent the similarities between cell pairs. Define a symmetric weight matrix W to quantify the edges. The elements in W represent the weights of the edges connecting cell i and cell j:
[0089]
[0090] where v i is the ith cell, N K (v i ) is v i The set of K nearest neighbors of . Therefore, graph regularization can be formulated as follows:
[0091]
[0092] The objective formula of non-negative matrix factorization is further rewritten as:
[0093]
[0094] Where D is a diagonal matrix and the elements in D are obtained by summing the rows or columns of W. i and f j Yes i and v j The low-dimensional representation of G. L is the Laplacian matrix of G, that is, L = DW.
[0095] In order to solve the independence problem of feature selection and cell clustering, projection matrix decomposition and non-negative matrix decomposition are combined to jointly learn feature selection and cell clustering. Combining formula (4) for feature selection and formula (8) for clustering, a joint learning model based on matrix decomposition method is constructed, and its objective function is defined as the following optimization problem:
[0096]
[0097] An iterative strategy is used to solve the target formula, optimizing one variable by fixing other variables until the termination criterion is reached. The performance of the joint learning model based on matrix decomposition for feature selection of single-cell data is shown in Figure 2. Figure 3 shown.
[0098] S3: Identify gene expression patterns and cell subpopulations: e.g. Figure 2 As shown in the figure, a joint learning model based on matrix decomposition is used to realize joint learning of multiple single-cell analysis tasks: the gene projection matrix is used to select representative genes and determine the gene function modules; the association matrix is used to calculate the importance of the modules and rank them; the coefficient matrix is used to cluster the single-cell data, and the cell type represented by each cluster is identified by comparing it with known cell marker genes;
[0099] S3.1 uses the cell projection matrix U obtained by the joint learning model based on matrix decomposition to select features in each column whose expression values are greater than the specified threshold as members of the module, thereby constructing a gene function module. i ,i=1,...,k, select the nodes whose weighted factor absolute value is greater than the threshold θ according to their z-score:
[0100]
[0101] in, is the mean of the row vectors of the matrix U, is the standard deviation of the row vector of U. Based on the above transformation, the condition z is satisfied ij >θ is taken as a member of the i-th module, and the threshold θ is set to the value corresponding to the highest point on the module similarity curve.
[0102] S3.2 quantifies the importance of the module through the association matrix S obtained by the joint learning model based on matrix decomposition. The element s of the association matrix S is ij As weights, u is used to reconstruct the matrix X i and v j A linear combination of ij The larger the value, the i and v jThe higher the similarity between the identified modules, the greater the element values of all combinations of the selected features in the matrix X. The correlation matrix S is a diagonal matrix whose non-zero elements are only on the diagonal. By evaluating the value of the matrix S, the importance of the identified modules can be determined.
[0103] S3.3 uses a joint learning model based on matrix decomposition to generate a coefficient matrix F, and assigns cells to cell clusters based on the maximum coefficient of matrix F. At the same time, combined with the expression level of marker genes in specific cell types, cell clusters are labeled as known cell types. This method can effectively identify and interpret cell clusters, and help to more fully understand the identity and function of cells. For those cell clusters that lack known marker gene expression information, they are assigned to the known cell type with the largest number of cells.
[0104] S4: Construction of an inference statistical model of intercellular communication: Based on the results of cell type identification, the scores of the communication signals transmitted by ligands and receptors between different types of cells, as well as the scores of the communication signals transmitted by downstream transcription factors and their target genes within cells are calculated respectively. By integrating intercellular and intracellular communication signals, the intercellular communication inference model based on single-cell transcriptome data is obtained;
[0105] S4.1 collects and integrates three layers of data information to construct the reference library DcjCommDB for intercellular communication as evidence for inferring interactions between cells, including: ligand-receptor (LR) interactions, transcription factor-target gene (TF-TG) interactions, and signal transduction from receptors to transcription factors. The DcjCommDB reference library collects 5597 human LR interactions, 18421 TGs regulated by 1537 TFs, and 411 TFs regulated by 1533 receptors. In addition, 5818 mouse LR interactions, 1457 TGs regulated by 1649 TFs, and 369 TFs regulated by 731 receptors were collected.
[0106] S4.2 Intercellular signaling fraction through L-LR interactions 2 norm to evaluate, that is LR k is a two-dimensional vector represented by the normalized expression values of the LR-interacting ligand and receptor:
[0107] LR k =(softmax(L i,k ),softmax(R j,k )) (11)
[0108] Among them, L i,k is the average expression value of ligand L in cell type i, R j,kis the average expression value of receptor R in cell type j. If the ligand L contains subunits and the expression values of the subunits are expressed, it is defined as the geometric mean of the expression values of all subunits: Likewise, the receptor R is defined as l g and r h are the expression values of g subunit and h subunit in the ligand and receptor, respectively.
[0109] S4.3 Combine the interactions between receptors, transcription factors and target genes to calculate the intracellular communication signal score S intra :
[0110]
[0111] Among them, TF PPR (t,r,p) represents the interaction between receptor r and transcription factor t in pathway p. TF activity (t,k) represents the activity of each transcription factor t in cell type k calculated according to Fisher's test:
[0112] TF activity (t,k)=1-P Fisher (t,k) (13)
[0113] S4.4 Combined intercellular communication fraction S inter and intracellular communication fraction S intra , calculate the cell interactions S between different cell types k , which is defined as:
[0114] S k =S inter ×S intra (14)
[0115] Predicting the performance of intercellular communication networks based on single-cell data Figure 6 shown.
[0116] In summary, the present invention infers the intercellular communication network by combining intercellular and intracellular communication signals. A joint learning model based on matrix decomposition is used to select features and detect gene function modules, thereby accurately identifying gene expression patterns. At the same time, cells are clustered and annotated to reveal the identity and function of cells. The inference statistical model of intercellular communication is used to integrate ligand-receptor pairs, transcription factors and their target genes to infer the intercellular communication network, which effectively solves the shortcomings of current recognition methods in terms of the lack of ability to predict potential connections between intracellular and extracellular communication.
[0117] The above disclosures are only several specific embodiments of the present invention. However, the embodiments of the present invention are not limited thereto. Any changes that can be conceived by those skilled in the art should fall within the protection scope of the present invention.
Claims
1. A method for inferring intercellular communication based on single-cell transcriptome data, characterized in that: The steps include: In the single-cell data, the genes with low expression levels are filtered, i.e., the genes expressed in less than a set percentage of cells, and then the genes with significant expression differences between different cells are selected, i.e., the genes with the highest standardized variance between different cells, i.e., the first set number of genes, and finally, regularization preprocessing is performed, and these single-cell data are recorded in the form of a matrix; The single-cell data matrix after regularization preprocessing is decomposed into a gene projection matrix, a cell projection matrix, and a correlation matrix, which respectively represent the expression of genes and cells in the projection space and the correlation between genes and cells; The cell projection matrix is further processed using a non-negative matrix decomposition method to obtain a coefficient matrix, which represents the cell expression in a low-dimensional space. The two decomposition methods are combined, and sparse constraints and graph regularity constraints are imposed on the cell projection matrix and the coefficient matrix, respectively, to propose a joint learning model based on matrix decomposition. A joint learning model based on matrix decomposition is used to achieve joint learning of multiple single-cell analysis tasks: representative genes are selected and gene function modules are determined using gene projection matrix; The association matrix was used to calculate the importance of gene function modules and rank the gene function modules; The single-cell data are clustered using the coefficient matrix and compared with known cell marker genes to identify the cell type represented by each cluster; Based on the results of cell type identification, the scores of the communication signals transmitted by ligands and receptors between different types of cells and the scores of the communication signals transmitted by downstream transcription factors and their target genes within cells are calculated respectively; By integrating intercellular and intracellular communication signals, an intercellular communication inference model based on single-cell transcriptome data is obtained.
2. A method for inferring intercellular communication based on single-cell transcriptome data according to claim 1, characterized in that: In the single-cell data, genes with low expression levels are filtered, and then genes with significant expression differences between different cells are selected, and finally regularization preprocessing is performed, including the following steps: 1.1) First, the genes with low expression in the single-cell data were filtered. These genes were expressed in less than 6% of the cells and were considered to be unable to provide useful information in the clustering process. 1.2) Secondly, for the filtered single-cell data, the principal component analysis method was used to select the top 2000 genes with the highest standardized variance between different cells, and log2 regularization was performed for downstream analysis: X i,j =log2(M i,j +1) (1) Among them, M i,j represents the expression level of the i-th gene in the j-th cell in the original single-cell data, X i,j Represents the expression level of the i-th gene in the j-th cell in the normalized single-cell data.
3. The method for inferring intercellular communication based on single-cell transcriptome data according to claim 1, characterized in that: The single-cell data matrix after regularization preprocessing is decomposed into a gene projection matrix, a cell projection matrix, and a correlation matrix, which respectively represent the expression of genes and cells in the projection space and the correlation between genes and cells; The cell projection matrix is further processed using the non-negative matrix decomposition method to obtain a coefficient matrix, which represents the cell expression in the low-dimensional space. The two decomposition methods are combined, and sparse constraints and graph regularization constraints are imposed on the cell projection matrix and the coefficient matrix, respectively. A joint learning model based on matrix decomposition is proposed, which includes the following steps: 2.1) Use the projection matrix decomposition method for feature selection to map the original data into a low-dimensional representation space; decompose the input single-cell gene expression matrix X into the gene projection matrix U, the cell projection matrix V and the correlation matrix S, so as to calculate the difference between the original features and the selected features: in, represents the Frobenius norm; X∈R m×n Represents the input single-cell gene expression matrix, containing m genes and n cells; gene projection matrix and the cell projection matrix Respectively represent the expression levels of genes and cells in the projection space, m is the number of genes, and n is the number of cells; is the correlation matrix, k1 is the dimension after decomposition; 2.2) In the process of feature selection, for the cell projection matrix V, a sparse constraint term L is introduced 2,1 Norm; L 2,1 The norm first calculates the L1 norm of the row vector, and then applies the L2 norm to the column vector: Among them, the data element v ij represents the element value of the i-th row and j-th column in the matrix V; the target formula of feature selection is rewritten as: 2.3) Cluster the cells using non-negative matrix factorization method; the non-negative matrix factorization method decomposes the matrix V into a base matrix B and a coefficient matrix F: in, and They are expressed as basis matrix and coefficient matrix respectively, m is the number of genes, n is the number of cells, and k2 is the dimension of dimensionality reduction; 2.4) Use graph regularization methods to constrain the non-negative matrix factorization process: construct a graph G to describe the proximity between cells in the coefficient matrix F, where the nodes on the graph represent cells and the edges on the graph represent the similarities between cell pairs; define a symmetric weight matrix W to quantify the edges, where the elements in W represent the weights of the edges connecting cell i and cell j: Among them, v i is the ith cell, N K (v i ) is v i The set of K nearest neighbors of ; then the formula of the graph regularization constraint is as follows: The objective formula of non-negative matrix factorization is further rewritten as: Where D is a diagonal matrix, and the elements in D are obtained by summing the rows or columns of W; f i and f j Yes i and v j The low-dimensional representation of G; L is the Laplacian matrix of G, that is, L = DW; Combining projection matrix decomposition and non-negative matrix decomposition, feature selection and cell clustering are jointly learned; combining formula (4) for feature selection and formula (8) for clustering, a joint learning model based on matrix decomposition method is constructed, and its objective function is defined as the following optimization problem: The objective formula is solved using an iterative strategy to optimize one variable by fixing the other variables until the termination criterion is reached.
4. The method for inferring intercellular communication based on single-cell transcriptome data according to claim 1, characterized in that: A joint learning model based on matrix decomposition is used to achieve joint learning of multiple single-cell analysis tasks: representative genes are selected and gene function modules are determined using the gene projection matrix; the importance of gene function modules is calculated and ranked using the association matrix; The single-cell data are clustered using the coefficient matrix and compared with known cell marker genes to identify the cell type represented by each cluster, including the following steps: 3.1) Using the cell projection matrix U obtained by the joint learning model based on matrix decomposition, the features with expression values greater than the specified threshold in each column are selected as members of the module, thereby constructing a gene function module; for each column vector u i ,i=1,...,k, select the nodes whose weighted factor absolute value is greater than the threshold θ according to their z-score: in, is the mean of the row vectors of the matrix U, is the standard deviation of the row vector of U; based on the above transformation, the condition z will be satisfied ij >θ's g(i) is regarded as a member of the i-th module, and the threshold θ is set to the value corresponding to the highest point on the module similarity curve; 3.2) The importance of the module is quantified by the association matrix S obtained by the joint learning model based on matrix decomposition; the element s of the association matrix S is ij As weights, u is used to reconstruct the matrix X i and v j The importance of the identified modules is determined by evaluating the value of the matrix S; 3.3) Generate the coefficient matrix F using a joint learning model based on matrix decomposition, and assign cells to cell clusters based on the maximum coefficient of the matrix F; at the same time, label the cell clusters as known cell types based on the expression levels of marker genes in specific cell types; for cell clusters that lack known marker gene expression information, assign them to the known cell type with the largest number of cells.
5. The method for inferring intercellular communication based on single-cell transcriptome data according to claim 1, characterized in that: Based on the results of cell type identification, the scores of the communication signals transmitted by ligands and receptors between different types of cells and the scores of the communication signals transmitted by downstream transcription factors and their target genes within cells are calculated respectively; by integrating intercellular and intracellular communication signals, an intercellular communication inference model based on single-cell transcriptome data is obtained, including the following steps: 4.1) Collect and integrate three layers of data information to build a reference library of cell-to-cell communication, DcjCommDB, as evidence for inferring cell-to-cell interactions, including: ligand-receptor interactions, transcription factor-target gene interactions, and signal transduction from receptors to transcription factors; 4.2) The intercellular signaling score is evaluated by the L2 norm of the ligand-receptor interaction, i.e. LR k is a two-dimensional vector representing the normalized expression values of the ligand and receptor of a ligand-receptor interaction: LR k =(softmax(L i,k ),softmax(R j,k ))(11) Among them, L i,k is the average expression value of ligand L in cell type i, R j,k is the average expression value of receptor R in cell type j; if ligand L contains subunits and represents the expression value of subunits, it is defined as the geometric mean of the expression values of all subunits: The receptor R is defined as l g and r h are the expression values of g subunit and h subunit in the ligand and receptor, respectively; 4.3) Combine the interactions between receptors, transcription factors and target genes to calculate the intracellular communication signal score S intra : Among them, TF PPR (t,r,p) represents the interaction between receptor r and transcription factor t in pathway p; TF activity (t,k) represents the activity of each transcription factor t in cell type k calculated according to Fisher's test: TF activity (t,k)=1-P Fisher (t,k) (13) 4.4) Combined intercellular communication fraction S inter and intracellular communication fraction S intra , calculate the cell interactions S between different cell types k , which is defined as: S k =S inter ×S intra (14)。 6. The method for inferring intercellular communication based on single-cell transcriptome data according to claim 5, characterized in that: The DcjCommDB reference library collects 5597 human ligand-receptor interactions, 18421 target genes regulated by 1537 transcription factors, and 411 transcription factors regulated by 1533 receptors. It also collects 5818 mouse ligand-receptor interactions, 1457 target genes regulated by 1649 transcription factors, and 369 transcription factors regulated by 731 receptors.
Citation Information
Cited By
Intercellular signal transduction inference method and device based on optimal transmission of knowledge graph, equipment and medium
CN119339796A
Method, device and equipment for identifying communication relationship between cells and storage medium
CN120808884A
Single cell signal path reconstruction method, system, equipment and medium
CN121565265A
