Cell analysis method and device integrating single cell transcriptome, block transcriptome and clinical phenotype information

Abstract

The invention discloses a cell analysis method and device for integrating single cell transcriptome, block transcriptome and clinical phenotype information, and relates to the field of cell analysis. The method comprises the following steps: reconstructing single cell transcriptome data based on components associated with difference characteristics of different clinical phenotypes, extracting dynamic components associated with clinical phenotype changes in single cell data by using a machine learning model, and completing single cell data reconstruction by using orthogonal projection transformation and inverse transformation in the machine learning model. Mathematical derivation proves that the used machine learning model (partial least squares classification analysis) can extract features associated with clinical phenotype changes from single cell or block transcriptome data on the basis of keeping a data topological structure to the maximum, and the method can judge clinical phenotype related cell subgroups with higher reliability. And irrelevant noise is effectively reduced.

Description

Technical Field

[0001] The present invention relates to the field of cell analysis, and in particular to a cell analysis method and device integrating single cell transcriptome, block transcriptome and clinical phenotype information. Background Art

[0002] Integrating clinical phenotypic features into the field of cell analysis remains a daunting challenge. Traditional studies of cell heterogeneity often rely on phenotypic information from single-cell data alone, ignoring the rich clinical data available in large-scale sequencing. Efforts to associate clinical information with single cells remain limited and lack noise robustness. Existing methods include Scissor [1] ,scDEAL [2] ,scAB [3] ,PACSI [4] ,and DEGAS [5] . The above methods all integrate multi-omics data by evaluating the similarity between single-cell and block transcriptomes, and then map the clinical phenotype information corresponding to the block transcriptome data to single cells. However, existing methods directly integrate single-cell transcriptome and block transcriptome data. Due to the mismatch between different types of data and various noise interferences, the results of existing methods cannot reliably estimate the clinical phenotype corresponding to each cell.

[0003] Therefore, technicians in this field are committed to developing a cell analysis method with less noise and capable of reliably estimating the characteristics of cell subpopulations, so as to gain a deeper understanding of the cellular / molecular microscopic characteristics related to disease progression and design more effective treatment plans.

[0004] References:

[0005] 1.Sun, D., et al., Identifying phenotype-associated subpopulations by integrating bulk and single-cell sequencing data. Nat Biotechnol, 2022.40(4):p.527-538.

[0006] 2. Chen, J., et al., Deep transfer learning of cancer drug responses by integrating bulk and single-cell RNA-seq data. Nat Commun, 2022.13(1):p.6494.

[0007] 3. Zhang, Q., S. Jin, and X. Zou, scAB detects multiresolution cell states with clinical significance by integrating single-cell genomics and bulksequencing data. Nucleic Acids Res, 2022.50(21):p.12112-12130.

[0008] 4. Liu, CH, et al., Identification of cell subpopulations associated with disease phenotypes from scRNA-seq data using PACSI. Bmc Biology, 2023.21(1).

[0009] 5. Johnson, TS, et al., Diagnostic Evidence GAuge of Single cells (DEGAS): a flexible deep transfer learning framework for prioritizing cell inrelation to disease. Genome Medicine, 2022.14(1). Summary of the invention

[0010] In view of the above-mentioned defects of the prior art, the technical problem to be solved by the present invention is to provide a cell analysis method with less noise and capable of reliably estimating the characteristics of cell subpopulations, so as to enable in-depth understanding of the cellular / molecular microscopic characteristics related to disease progression and further design more effective treatment plans.

[0011] To achieve the above objectives, the present invention provides a cell analysis method integrating single-cell transcriptome, block transcriptome and clinical phenotype information, characterized in that the method integrates single-cell transcriptome data and block transcriptome data, and the block transcriptome data includes matching clinical phenotype meta-information.

[0012] In a preferred embodiment of the present invention, the clinical phenotype meta-information includes survival and treatment responsiveness.

[0013] In another preferred embodiment of the present invention, the method is specifically:

[0014] Extract shared biological features across single-cell and bulk transcriptome data, representing single-cell transcriptome data as X S ={x S1 , x S2 , ..., x Sn1}, and denote the block transcriptome data as X B ={x B1 , x B2 , ..., x Bn2},in Represents cell S from single-cell transcriptome data i and sample B from block transcriptome data j ,Here, N represents the number of shared genes, n1 and n2 represent the number of cells and samples, respectively;

[0015] Construct a comparison function, namely equation 1:

[0016]

[0017] In Equation 1, the transformation matrix C = {c1, c2, ..., c n} and P = {p1, p2, ..., p n};

[0018] Reduce X by orthogonal transformation C S The dimension of , and keeping the squared reconstruction error as small as possible, is achieved by the following equation 2:

[0019]

[0020] Among them, ‖.‖ represents the L2 norm;

[0021] After a slight modification of Equation 1, we obtain Equation 3, which is the cost function of the partial least squares PLS model:

[0022]

[0023] By nonlinear iterative partial least squares, i.e., using NIPALS, the first components c1 and p1 of the transformation matrix are calculated as Equation 4:

[0024]

[0025] w=w / ||w||

[0026] t=X S w / w T w

[0027]

[0028] Repeat the process of equations 1-3 n times to estimate C and P, let X′S =X S CC -1 and X′ B =X B PP -1 , X′ S and X′ B are reconstructed single-cell and block transcriptome data containing only the desired features.

[0029] In another preferred embodiment of the present invention, partial least squares discriminant analysis (PLSDA) is used to transform the single cell data X′ S and X′ B Projected into another feature space, the new feature space is composed of dynamic features related to the discrimination of different clinical phenotypes, namely, pseudo-dynamic space, based on X′ B and clinical metadata to construct a pseudo-dynamic spatial projection model, which is then applied to X′ S , the specific steps are as follows:

[0030] Assume that the meta information of each sample is related to its X′ B There is a one-to-one correspondence. For binary phenotypic conditions, a virtual matrix Y of size n2×2 is constructed, which contains only binary values. The two columns represent the phenotypic conditions to which each sample belongs, represented by (1,0) or (0,1). The covariance-based contrast function is optimized by using the NIPALS method, as shown in the following equation 5:

[0031]

[0032] A new transformation matrix P′ is obtained, and the space formed by P′ represents the features related to the discrimination of the two phenotypic conditions;

[0033] The single cell data is passed through X′ B P′ is projected into the pseudo-dynamic space, as shown in Equation 6:

[0034] X″ S =X′ S P′P′ -1 .

[0035] In another preferred embodiment of the present invention, the SeuratR package is used to analyze the cell X″ S Perform clustering.

[0036] In another preferred embodiment of the present invention, the specific method of clustering is as follows:

[0037] Use Seurat's FindAllMarkers function to obtain specific highly expressed genes in each cluster;

[0038] Calculate the Pearson correlation matrix To confirm the reconstruction of X″ S In the data, all components irrelevant to the different phenotypic distinctions have been eliminated, and the highly expressed markers identified from each cell cluster represent the contribution of these cells to the phenotypic condition distinction, where r ij represents the similarity between each cell in the cluster and a block sample on the identified marker gene, and k represents the number of cells in the cluster.

[0039] In another preferred embodiment of the present invention, OPLSDA, orthogonal partial least squares discriminant analysis, is used to test the significance of similarity. The specific method is as follows:

[0040] Search for common features between the cell-sample correlation matrix R and a dummy matrix Y, which indicates the phenotype to which each sample belongs, and query the first PLS component to extract the desired information;

[0041] The OPLSDA model calculation between R and Y is shown in Equation 7:

[0042] p″1=R T t / t T t via max Cov(RP″,YQ′)

[0043] w o =p″1-(w T p″1 / w T w)w

[0044] w o =w o / ||w o ||

[0045]

[0046] p o =R T t / t T t

[0047]

[0048] max Cov(R OPLS P″′,YQ′).

[0049] In another preferred embodiment of the present invention, in order to identify the association between a cell and a phenotype, the correlation between cell i and the first OPLSDA score t OPLS，1 The correlation between them is denoted as r i-OPLS :

[0050] Significant r i-OPLS , that is, a p value < 0.05 indicates that cell i is more similar to a certain phenotype;

[0051] If rij(j∈phenotype A)>rij(j∈phenotype B) ,and Then cell i is associated with phenotype A, where A and B refer to the relative situation of a clinical phenotype, such as A represents a good prognosis and B represents a poor prognosis.

[0052] The present invention also provides a cell analysis device that integrates single-cell transcriptome, block transcriptome and clinical phenotype information, characterized in that the analysis device executes the aforementioned method.

[0053] In a preferred embodiment of the present invention, the device includes a feature extraction module for extracting shared biological features across single-cell and block transcriptome data; a projection module for projecting single-cell data into another feature space; a clustering module for clustering the projected data; and a discriminant analysis module for testing the significance of similarities and identifying associations between cells and phenotypes.

[0054] Technical Effects

[0055] The present invention identifies cell populations associated with clinical phenotypes by integrating multi-omics data. Unlike existing methods and devices, the methods and devices of the present invention reconstruct single-cell transcriptome data based on components associated with differential features of different clinical phenotypes, and use machine learning models to extract dynamic components related to clinical phenotype changes in single-cell data, wherein clinical phenotype features come from block transcriptome data, and the machine learning model trained from block transcriptome data is applied to single-cell transcriptome data to achieve the above purpose; secondly, the orthogonal projection transformation in the machine learning model and its inverse transformation are used to complete the reconstruction of single-cell data. Mathematical derivation proves that the machine learning model (partial least squares classification analysis) used can extract features associated with clinical phenotype changes from single-cell or block transcriptome data on the basis of maximizing the maintenance of the data topology. The present invention determines clinical phenotype-related cell subpopulations with higher reliability, effectively reduces irrelevant noise, and shows higher reliability in determining clinical phenotype-related cell subpopulations. It can help to gain a deeper understanding of the cell / molecular microscopic features related to disease progression, and then design more effective treatment plans.

[0056] The concept, specific structure and technical effects of the present invention will be further described below in conjunction with the accompanying drawings to fully understand the purpose, characteristics and effects of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 Schematic diagram of a cell analysis method integrating single cell transcriptome, block transcriptome and clinical phenotype information in a preferred embodiment of the present invention DETAILED DESCRIPTION

[0058] The following describes several preferred embodiments of the present invention with reference to the drawings in the specification, so that the technical content is clearer and easier to understand. The present invention can be embodied in many different forms of embodiments, and the protection scope of the present invention is not limited to the embodiments mentioned in the text.

[0059] like Figure 1 As shown, this method is suitable for the integrated analysis of single-cell data and block transcriptome data, where the block transcriptome data needs to have matching clinical phenotype meta-information, such as survival, treatment responsiveness, etc.

[0060] Given the bulk transcriptome data as well as the single-cell transcriptome data, we extracted the shared biological features across the single-cell and bulk transcriptome data. The single-cell transcriptome data is represented as X S ={x S1 , x S2 , ..., x Sn1}, and denote the block transcriptome data as X B ={x B1 , x B2 , ..., x Bn2},in Represents cell S from single-cell transcriptome data i and sample B from block transcriptome data j Here, N represents the number of shared genes, and n1 and n2 represent the number of cells and samples, respectively.

[0061] Since these two types of data are intended to describe the same problem, it is reasonable to assume that there are low-dimensional feature spaces in which the projections of single-cell RNA-seq and RNA-seq data are similar. At the same time, the configuration of the original data can be preserved to the greatest extent in the feature space. To meet the above requirements, we construct a contrast function Equation 1:

[0062]

[0063] We hope to reduce X by an orthogonal transformation C S dimensionality, and we want to keep the squared reconstruction error as small as possible, see Equation 2:

[0064]

[0065] ‖.‖ represents the L2 norm.

[0066] Therefore, the first two terms in Equation 1 ensure that the constructed transformation matrix C = {c1, c2, ..., c n} and P = {p1, p2, ..., p n}Maximum retention of X S and X BThe last item Corr(X S C, X B P) ensures that the projected X S and X B has the largest correlation. After slightly modifying Equation 1, we obtain Equation 3:

[0067]

[0068] This is the cost function for the partial least squares (PLS) model.

[0069] A widely used solution to Equation 3 is Nonlinear Iterative Partial Least Squares (NIPALS). Using NIPALS, the first components c1 and p1 of the transformation matrix can be calculated as Equation 4, as follows:

[0070]

[0071] w=w / ||w||

[0072] t=X S w / w T w

[0073]

[0074]

[0075] Repeat the above process n times to estimate C and P.

[0076] Let X′ S =X S CC -1 and X′ B =X B PP -1 , these are reconstructed single-cell and block transcriptome data containing only the desired features, which are X S and X B , and are highly correlated with each other.

[0077] Once the latent feature X′ is extracted S and X′ B , partial least squares discriminant analysis (PLSDA) is applied to project the single-cell data into another feature space. The new feature space should be composed of dynamic features related to the discrimination of different clinical phenotypes, so it is called pseudo-dynamic space. The pseudo-dynamic space is intended to represent the single-cell resolution gene expression dynamics between different phenotypes specified by the clinical metadata of the block transcriptome dataset. In the present invention, based on X′ B A pseudo-dynamic spatial projection model was constructed based on clinical metadata, and then applied to X′ S .

[0078] It is reasonable to assume that the meta-information of each sample (patient or animal) is consistent with its X′ B There is a one-to-one correspondence.

[0079] Taking the binary phenotypic condition as an example, a virtual matrix Y of size n2×2 is constructed, which contains only binary values. The two columns represent the phenotypic condition to which each sample belongs, represented by (1,0) or (0,1). When optimizing the covariance-based contrast function using the NIPALS method, see Equation 5:

[0080]

[0081] A new transformation matrix P′ is obtained, and the space spanned by P′ represents the features related to the discrimination of the two phenotypic conditions.

[0082] Single-cell data, i.e., RNA-seq data, can be obtained through X′ B P′ is projected into the pseudo-dynamic space as shown in Equation 6:

[0083] X″ S =X′ S P′P′ -1

[0084] Seurat R package was used to analyze cells X″ S Perform clustering:

[0085] Use Seurat's 'FindAllMarkers' function to obtain the specific highly expressed genes in each cluster. In the reconstructed X" S In the data, all components not related to the different phenotypic distinctions have been eliminated, and the highly expressed markers identified from each cell cluster represent the contribution of these cells to the phenotypic condition distinction. To confirm this, the Pearson correlation matrix was calculated. where r ij represents the similarity between each cell in the cluster and a block sample on the identified marker gene, and k represents the number of cells in the cluster.

[0086] A reasonable assumption is that if a cell cluster has a higher correlation with block samples of a certain phenotype and a lower correlation with block samples of another phenotype, then this cell cluster may be associated with this phenotype. To test the significance of the similarity, OPLSDA (Orthogonal Partial Least Squares Discriminant Analysis), an extended version of PLSDA, is used. Here, the common features of the cell-sample correlation matrix R and a dummy matrix Y (indicating the phenotype to which each sample belongs) are searched. Compared with the standard PLSDA, OPLSDA includes an orthogonal filtering step to compress all the discriminant features into the first partial least squares (PLS) component, which can extract the required information by querying the first PLS component alone. The OPLSDA model between R and Y is calculated as shown in Equation 7:

[0087] p″1=R T t / t T t via maxCov(RP″,YQ′)

[0088] w o =p″1-(w T p″1 / w T w)w

[0089] w o =w o / ||w o ||

[0090]

[0091] p o =R T t / t T t

[0092]

[0093] max Cov(R OPLS P″′,YQ′)

[0094] To identify cell-phenotype associations, cell i is evaluated for its relationship to the first OPLSDA score t OPLS，1 The correlation between them is denoted as r i-OPLS . Significant r i-OPLS (p value < 0.05) indicates that cell i is more similar to a certain phenotype. ij(j∈phenotype A)>rij(j∈phenotype B ),and Then cell i is associated with phenotype A, where A and B refer to the relative situation of a clinical phenotype, such as A represents a good prognosis and B represents a poor prognosis.

[0095] The present invention also includes a cell analysis device that integrates single-cell transcriptome, block transcriptome and clinical phenotype information, which can execute the above method. The device includes a feature extraction module for extracting shared biological features across single-cell and block transcriptome data; a projection module for projecting single-cell data into another feature space; a clustering module for clustering the projected data; and a discriminant analysis module for testing the significance of similarity and identifying the association between cells and phenotypes.

[0096] The preferred specific embodiments of the present invention are described in detail above. It should be understood that ordinary technicians in the field can make many modifications and changes based on the concept of the present invention without creative work. Therefore, all technical solutions that can be obtained by technicians in the technical field based on the concept of the present invention through logical analysis, reasoning or limited experiments on the basis of the prior art should be within the scope of protection determined by the claims.

Claims

1. A cell analysis method integrating single cell transcriptome, bulk transcriptome and clinical phenotype information, characterized in that: The method integrates single-cell transcriptome data and bulk transcriptome data including matched clinical phenotype meta-information.

2. The method according to claim 1, characterized in that The clinical phenotype meta-information includes survival and treatment responsiveness.

3. The method according to claim 2, characterized in that The method is specifically as follows: Extract shared biological features across single-cell and bulk transcriptome data, representing single-cell transcriptome data as X S ={x S1 , x S2 , ..., x Sn1 }, and denote the block transcriptome data as X B ={x B1 , x B2 , ..., x Bn2 },in Represents cell S from single-cell transcriptome data i and sample B from block transcriptome data j ,Here, N represents the number of shared genes, n1 and n2 represent the number of cells and samples, respectively; Construct a comparison function, namely equation 1: In Equation 1, the transformation matrix c = {c1, c2, ..., c n } and P = {p1, p2, ..., p n }; Reduce X by orthogonal transformation C S The dimension of , and keeping the squared reconstruction error as small as possible, is achieved by the following equation 2: Among them, ‖.‖ represents the L2 norm; After a slight modification of Equation 1, we obtain Equation 3, which is the cost function of the partial least squares PLS model: By nonlinear iterative partial least squares, i.e., using NIPALS, the first components c1 and p1 of the transformation matrix are calculated as Equation 4: w=w / ||w|| t=X S w / w T w Repeat the process of equations 1-3 n times to estimate C and P, let X′ S =X S CC -1 and X′ B =X B PP -1 , X′ S and X′ B are reconstructed single-cell and block transcriptome data containing only the desired features.

4. The method according to claim 3, characterized in that Partial least squares discriminant analysis (PLSDA) is used to transform the single cell data X′ S and X′ B Projected into another feature space, the new feature space is composed of dynamic features related to the discrimination of different clinical phenotypes, namely, pseudo-dynamic space, based on X′ B and clinical metadata to construct a pseudo-dynamic spatial projection model, which is then applied to X′ S , the specific steps are as follows: Assume that the meta information of each sample is related to its X′ B There is a one-to-one correspondence. For binary phenotypic conditions, a virtual matrix Y of size n2×2 is constructed, which contains only binary values. The two columns represent the phenotypic conditions to which each sample belongs, represented by (1,0) or (0,1). The covariance-based contrast function is optimized by using the NIPALS method, as shown in the following equation 5: A new transformation matrix P′ is obtained, and the space formed by P′ represents the features related to the discrimination of the two phenotypic conditions; The single-cell data are projected into the pseudo-dynamic space as shown in Equation 6: X″ S =X′ S P′P′ -1 。 5. The method according to claim 4, characterized in that SeuratR package was used to analyze cell X″ S Perform clustering.

6. The method according to claim 5, characterized in that The specific method of clustering is as follows: Use Seurat's FindAllMarkers function to obtain specific highly expressed genes in each cluster; Calculate the Pearson correlation matrix To confirm the reconstruction of X″ S In the data, all components irrelevant to the different phenotypic distinctions have been eliminated, and the highly expressed markers identified from each cell cluster represent the contribution of these cells to the phenotypic condition distinction, where r ij represents the similarity between each cell in the cluster and a block sample on the identified marker gene, and k represents the number of cells in the cluster.

7. The method according to claim 6, characterized in that OPLSDA, orthogonal partial least squares discriminant analysis, is used to test the significance of similarity. The specific method is as follows: Search for common features between the cell-sample correlation matrix R and a dummy matrix Y, which indicates the phenotype to which each sample belongs, and query the first PLS component to extract the desired information; The OPLSDA model calculation between R and Y is shown in Equation 7: p″1=R T t / t T t via maxCov(PR″,YQ′) w o =p″1-(w T p″1 / w T w)w In o =in o / ||in o || p o =R T t / t T t maxCov(R OPLS P″′,YQ′)。 8. The method according to claim 7, characterized in that To identify cell-phenotype associations, cell i is evaluated for its relationship to the first OPLSDA score t OPLS，1 The correlation between them is denoted as r i-OPLS : Significant r i-OPLS , that is, a p value < 0.05 indicates that cell i is more similar to a certain phenotype; If r ij(j∈phenotype A) >r ij(j∈phenotype B) ,and Then cell i is associated with phenotype A, where A and B refer to the relative situation of a clinical phenotype, such as A represents a good prognosis and B represents a poor prognosis.

9. A cell analysis device integrating single cell transcriptome, bulk transcriptome and clinical phenotype information, characterized in that: The analysis device performs the method according to any one of claims 1-8.

10. The cell analysis device according to claim 9, characterized in that: The apparatus includes a feature extraction module for extracting shared biological features across single-cell and block transcriptome data; A projection module is used to project the single-cell data into another feature space; a clustering module is used to perform clustering on the projected data; Discriminant analysis module, used to test the significance of similarities and identify associations between cells and phenotypes.

Images