Cell group division method based on overburden remodeling coordination

Abstract

A cell group division method based on overburden remodeling coordination belongs to the field of cell group division in single-cell RNA sequencing data analysis, and comprises the following steps: firstly, constructing cell multi-fractal dimension characteristics for single-cell RNA sequencing data so as to better utilize different fractal dimension characteristic information; and then multi-fractal-dimension low-dimensional representation is extracted. Secondly, extracting multi-fractal-dimension fusion representation, multi-fractal-dimension remodeling representation and cross-fractal-dimension prediction representation by adopting an obstacle covering mechanism so as to improve the robustness of fractal-dimension remodeling; and optimizing the initial division loss to obtain the initial division of the cells. Then, a confidence weight is obtained based on the remodeling error, the division entropy, and the suspension to reduce the negative effects of the low-quality fractal dimension data. And the consistency and coordination of cell group division are ensured by minimizing the division loss target, and a final cell group division result is obtained. Compared with other methods, the cell group division accuracy is higher, and the performance is more stable.

Description

TECHNICAL FIELD

[0001] The application belongs to the field of single-cell RNA sequencing data analysis, and particularly relates to a cell cluster division method based on coverage barrier remodeling coordination. BACKGROUND

[0002] The rapid development of single-cell RNA sequencing (scRNA-seq) technology enables researchers to obtain gene expression profiles at the single-cell level, thereby deeply analyzing the transcriptional heterogeneity among cells. Compared with traditional bulk sequencing methods, scRNA-seq not only breaks through the limitations of traditional bulk cell sequencing technology, but also reveals the gene expression differences of different cells in the cell population, providing a powerful tool for studying cell function, developmental trajectory, disease mechanism, and molecular regulatory network. In the downstream processing of scRNA-seq data, cell cluster division is a key step. The purpose is to divide cells with similar gene expression patterns into the same cluster, helping researchers to analyze the biological function of cells, reveal potential cell subpopulations, and explore their regulatory patterns. However, scRNA-seq data itself has some significant characteristics that pose challenges to technical applications: 1) high dimensionality of data itself: increasing the complexity of computational processing; 2) significant sparsity: a large number of genes in the gene expression profile show zero expression, affecting the accurate modeling of expression patterns; 3) noise interference: due to factors such as batch effects, biological changes, and technical noise, there is noise in the data that interferes with the reliability of cell cluster division. Existing traditional methods are mostly based on a single analysis strategy, which cannot fully capture the multidimensional characteristics of the data, and it is also difficult to effectively deal with the above challenges, resulting in unsatisfactory cell cluster division. Therefore, it is urgent to propose a new technical solution that can effectively deal with the high dimensionality and high sparsity characteristics of scRNA-seq data to achieve accurate cell cluster division. SUMMARY

[0003] To solve the above problems, the application provides a cell cluster division method based on coverage barrier remodeling coordination, which comprises the following steps:

[0004] Obtain a single-cell RNA sequencing data set, remove low-expression-value genes from the data set to obtain a single-cell RNA gene expression matrix X, and use PCA, DNMF, KPCA, etc. Three kinds of feature encoders to construct cell multi-dimensional feature X (v) v = 1, 2, 3;

[0005] The cell multi-dimensional feature X (v) is calculated using residual connection through a nonlinear layer, group normalization, and GELU activation function to obtain a multi-dimensional low-dimensional representation H (v) ;

[0006] Set a coverage barrier set M, which is composed of X(v) Indexing the low-quality data, using multi-dimensional fusion representation H (f) ;

[0007] Using multi-dimensional fusion representation H (v) Reconstructing the multi-dimensional representation of the low-quality data using multi-dimensional fusion representation H And constructing reconstruction loss L rec Using multi-dimensional fusion representation H (f) Generating cross-dimensional prediction representation And constructing prediction loss L pres ;

[0008] Optimizing the initial partition loss L p1 Using H (v) and H (f) Using the cell partitioner to obtain the initial cell partition Q (v) ;

[0009] Calculating reconstruction error, partition entropy, suspension and confidence weight for the multi-dimensional representation of the cells;

[0010] Constructing prediction loss L pres , partition coordination loss L par and partition entropy loss L paren , optimizing the partition loss L p2 Training the model to convergence;

[0011] Calculating the final cell partition from the initial cell partition Q (v) of the converged model, and obtaining the accuracy of the cell cluster partition by combining the true labels of the cells in the data set.

[0012] Further, the multi-dimensional features X (v) of the i-th cell are defined as follows:

[0013]

[0014] Where n is the total number of cells, d v is the feature dimension.

[0015] Using PCA on X (v) , calculate the expression median deviation of each gene, if the number of genes in the v-th dimension of the i-th cell that exceed the expression median deviation threshold exceeds the set threshold, record the index value pair of the dimension and the cell in the cover set M, M is defined as follows:

[0016]

[0017] If (i, v) ∈ M, then The data features will be obscured, thus reducing the negative impact of noise or abnormal gene expression values ​​on features in single-cell RNA sequencing data.

[0018] Feature extraction of uncovered cells and multidimensional fusion representation of H (f) The definition is as follows:

[0019]

[0020] in, and H respectively (f) and H (v) The corresponding low-dimensional representation of the i-th cell, F() will Perform an outer product operation to map the result to a high-dimensional tensor, then use an attention mechanism to obtain... [,] combines the input components into a single variable. Composition H (f) ;

[0021] Furthermore, the multidimensional reconstruction of cells represents The definition is as follows:

[0022]

[0023] Among them, D (v) The input is processed through two non-linear layers, an average pooling layer, and the ReLU activation function to extract features.

[0024] Reshaping loss L rec The definition is as follows:

[0025]

[0026] Where V is the total fractal dimension, ⊙ is the element-wise multiplication operation, and γ is the total number of elements. (v) For dimension weights, Let be a learnable parameter vector, where ∈ represents non-negative parameters;

[0027] Cross-dimensional prediction representation The definition is as follows:

[0028]

[0029] Predicted loss L pres The definition is as follows:

[0030]

[0031] Furthermore, to minimize L p1 To achieve the goal, the initial cell division Q is obtained through a cell divider. (v) Initial division loss L p1 The definition is as follows:

[0032] L p1 = L rec + λ1L pres

[0033] where λ1is the balancing parameter, and the cell partitioner is trained to reduce the remodeling and prediction loss;

[0034] The implementation of the cell partitioner is defined as follows:

[0035]

[0036] where K is the total number of cell classes, is the case that the ith cell in the vth dimension belongs to each class, softmax() is the normalization activation function, linear() is the linear function, τ is the controllable temperature index, f φ () is the inference network;

[0037] Further, the remodeling error of the ith cell in the vth dimension is defined as follows:

[0038]

[0039] The partition entropy of the ith cell in the vth dimension is defined as follows:

[0040]

[0041] where, is the kth column of , indicating the possibility that the ith cell in the vth dimension belongs to class k;

[0042] The suspension u of the partition of the ith cell in the vth dimension i,v is defined as follows:

[0043]

[0044] where α is the weight coefficient, and norm() is the normalization operation;

[0045] The confidence weight of the partition of the ith cell in the vth dimension is defined as follows:

[0046]

[0047] where β is the temperature scaling hyperparameter;

[0048] Further, the partition coordination loss L par is defined as follows:

[0049]

[0050] where || ||2 is the L2 norm;

[0051] partition entropy loss L paren is defined as follows:

[0052]

[0053] partition loss L p2 is defined as follows:

[0054] L p2 = L p1 + λ2L par + λ3L paren

[0055] where λ2 and λ3 are balance parameters, and L p1 is constructed. p2 On the basis of L p1 , L p2 is minimized by Adam to reduce the partition loss.

[0056] Further, the calculation method of the final cell partition is as follows:

[0057]

[0058] where, is the comprehensive partition of the i-th cell in all sub-dimensions, q i is the final class label of the i-th cell, which is the class label corresponding to the maximum value in

[0059] The application provides a cell class group partition method based on coverage barrier remodeling coordination, which has the following advantages:

[0060] (1) The method constructs multiple sub-dimension features of cells and covers barriers for low-quality data, uses fusion representation to generate cross-sub-dimension prediction to enhance the robustness of the method to noise and abnormal gene expression in single-cell RNA sequencing data;

[0061] (2) The method continuously optimizes cell partition by minimizing remodeling loss, prediction loss, partition coordination loss and partition entropy loss to achieve consistency of multi-sub-dimension partition results;

[0062] (3) The method automatically reduces the influence of noise and low-quality sub-dimension data by confidence weight, and improves the accuracy of partitioning cell class groups. BRIEF DESCRIPTION OF DRAWINGS

[0063] ​To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0064] Figure 1 This is a flowchart of a cell grouping method based on barrier remodeling coordination for single-cell RNA sequencing data provided by the present invention. Detailed Implementation

[0065] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. It should be understood that these descriptions are merely exemplary and not intended to limit the scope of the invention. Furthermore, descriptions of well-known structures and techniques are omitted in the following description to avoid unnecessarily obscuring the concept of the invention.

[0066] Exemplary methods

[0067] like Figure 1 This invention provides a cell group classification method based on barrier remodeling coordination, the method steps are as follows:

[0068] Step S110: Obtain the single-cell RNA sequencing dataset, remove genes with low expression values ​​from the dataset to obtain the single-cell RNA gene expression matrix X, and construct the cell multidimensional feature X using three feature encoders: PCA, DNMF, and KPCA. (v) v = 1, 2, 3, Cellular multidimensional characteristics X (v) The multidimensional features of the i-th cell are defined as follows:

[0069]

[0070] Where n is the total number of cells, d v For feature dimensions;

[0071] Step S120: Analyze the multidimensional features of the cell X (v) The multidimensional low-dimensional representation H is obtained by using residual connections, nonlinear layers, group normalization, and GELU activation function. (v) ;

[0072] Step S130: For X (v) After using PCA, the median expression deviation is calculated for each gene. If the number of genes exceeding the median expression deviation threshold in the i-th cell of the v-th fractal exceeds a set threshold, the index pairs of the fractal and the cell are recorded in the barrier set M, which is defined as follows:

[0073]

[0074] If (i, v)∈M, then The covered obstacles, whose data features are not available; extract the multi-fractal fusion representation of the cells of the uncovered obstacles (f) , defined as follows:

[0075]

[0076] where, and are the low-dimensional representations of the i-th cell in H (f) and H (v) , respectively, and F() maps the outer product operation to a high-dimensional tensor using an attention mechanism to obtain , which combines the input components into a whole variable, and into H (f) ;

[0077] Step S140: Extract the multi-fractal remodeling representation of the cells of the multi-fractal low-dimensional representation H (v) in which the fractal features of the uncovered obstacles are extracted defined as follows:

[0078]

[0079] where, D (v) extracts features from the input through two nonlinear layers, an average pooling layer, and an activation function ReLU;

[0080] The remodeling loss L rec is defined as follows:

[0081]

[0082] where V is the total number of fractal dimensions, is the element-wise multiplication operation, γ (v) is the dimension weight, is a learnable parameter vector, and ∈ is a non-negative parameter; the multi-fractal fusion representation H (f) is used to generate a cross-fractal prediction representation defined as follows:

[0083]

[0084]

[0085] The prediction loss L pres is defined as follows:

[0086]

[0087] Step S150: Initial partition loss L p1 is defined as follows:

[0088] L p1 = L rec + λ1L pres

[0089] where λ1 is a balance parameter;

[0090] minimize L p1 to obtain the initial partition Q of cells by a cell partitioner (v) The implementation of the cell partitioner is defined as follows:

[0091]

[0092] where K is the total number of cell classes, is the case of the i-th cell belonging to each class in the v-th dimension, softmax() is a normalized activation function, linear() is a linear function, τ is a controllable temperature index, f φ () is an inference network;

[0093] Step S160: Calculate the remodeling error, partition entropy, suspension and confidence weight of the multi-dimensional representation of the cell; the remodeling error of the i-th cell in the v-th dimension is defined as follows:

[0094]

[0095] The partition entropy of the i-th cell in the v-th dimension is defined as follows:

[0096]

[0097] where is the k-th column of , indicating the possibility of the i-th cell in the v-th dimension belonging to class k;

[0098] The suspension u of the i-th cell partition in the v-th dimension i,v is defined as follows:

[0099]

[0100] where α is a weight coefficient, and norm() is a normalization operation;

[0101] The confidence weight of the i-th cell partition in the v-th dimension is defined as follows:

[0102]

[0103] wherein β is a hotness scaling hyper-parameter;

[0104] Step S170: dividing coordination loss L par which is defined as follows:

[0105]

[0106] dividing entropy loss L paren which is defined as follows:

[0107]

[0108] constructing dividing loss L p2 which is defined as follows:

[0109] L p2 = L p1 + λ2L par + λ3L paren

[0110] wherein λ2and λ3are balance parameters, and L p2 training the model to convergence;

[0111] Step S180: obtaining cell initial division Q (v) to obtain the final cell division, the calculation method is as follows:

[0112]

[0113] wherein, is the comprehensive division of the i-th cell in all dimensions, q i is the final class label of the i-th cell, which is the class label corresponding to the maximum value in The accuracy of cell class division is obtained by combining the real labels of cells in the data set.

[0114] Through the embodiment, first, the multi-dimensional features of cells are constructed, and then the multi-dimensional low-dimensional representation is extracted to highlight the key features. Secondly, the low-quality gene data coverage mechanism is introduced, and then the multi-dimensional remodeling representation of cells is extracted for the un-covered dimensional features. The features of the covered dimensions are predicted by the multi-dimensional fusion representation, different dimensional information is predicted, further, the dimensional division uncertainty is defined based on the confidence weight and the division coordination entropy, which is used for division prediction and class coordination to reduce the negative impact of low-quality dimensional data, and finally the cell class division is obtained. The accuracy of the division is obtained by combining the real class labels of cells in the data set.

[0115] Further explanation, assuming that a single cell RNA sequencing data set is divided into cell classes according to the present embodiment, a class division result with an accuracy higher than most mainstream methods can be obtained.

[0116] DETAILED DESCRIPTION

[0117] The embodiment adopts a published single-cell RNA sequencing dataset. Details of the dataset are described as follows:

[0118] The Deng dataset is from the National Center for Biotechnology Information (https: / / www.ncbi.nlm.nih.gov / ). The dataset is composed of single cells of human cortex in development, including 268 cells, 22431 genes, and corresponding to 7 cell clusters.

[0119] In order to verify the superiority of the embodiment (Ours), the embodiment is compared with other four baseline methods for cell cluster partitioning of single-cell RNA sequencing data, including t-SNE, SC3, scCAN, SCEA and the like. The clustering accuracy (ACC) of these methods on the above published dataset will be compared. The specific data comparison is shown in Table 1.

[0120] Table 1. Cell cluster partitioning accuracy (%) of Deng dataset

[0121] t-SNE SC3 scCAN SCEA Ours 62.2 63.0 68.1 66.7 72.9

[0122] From the above data comparison, it can be clearly seen that Ours achieves the best performance and significantly improves the accuracy of cell cluster partitioning for single-cell RNA sequencing data. The quantitative results fully illustrate the superiority of Ours, because Ours constructs multi-dimensional features and performs remodeling representation to highlight key features, implements an occlusion mechanism to reduce the influence of low-quality data on modeling. At the same time, uncertainty perception is performed based on remodeling error and partitioning entropy to avoid noise interference, so as to accurately model the cell gene pattern and further improve the accuracy of cell cluster partitioning.

[0123] The embodiment proposes a cell cluster partitioning method based on occlusion remodeling coordination, which is used for cell cluster partitioning of single-cell RNA sequencing data. Multi-dimensional cell features are constructed for single-cell RNA sequencing data, and then multi-dimensional low-dimensional representation is extracted. Then, an occlusion mechanism is performed, and multi-dimensional fusion representation, multi-dimensional remodeling representation and cross-dimensional prediction representation are extracted for remodeling and different dimensional prediction. Confidence weights are obtained based on remodeling error, partitioning entropy and suspension to suppress the influence of low-quality data and improve the quality of partitioning prediction. The overall target loss is set, and the loss is constantly minimized to optimize cell cluster partitioning, and the final cell cluster partitioning result is obtained.

[0124] The experimental results on the public single-cell RNA sequencing dataset Deng show that the present embodiment has higher accuracy and better superiority in cell cluster division than other methods.

[0125] It should be understood that the above specific embodiments of the present application are merely used for illustrative or explanatory purposes of the principles of the present application, and do not constitute a limitation on the present application. Therefore, any modification, equivalent replacement, improvement, etc. made without departing from the spirit and scope of the present application shall be included in the protection scope of the present application. In addition, the appended claims of the present application are intended to cover all variations and modifications falling within the scope and boundary of the appended claims, or the equivalent forms of such scope and boundary.

Claims

1.A method for cell type classification based on overlay reshaping coordination, characterized in that, The method comprises the steps of: S110. Obtain a single-cell RNA sequencing data set, remove genes with low expression values from the data set to obtain a single-cell RNA gene expression matrix X, and use three characteristic encoders such as PCA, DNMF, and KPCA to construct cell multi-dimensional features X (v) v = 1, 2, 3; S120. Obtain the multi-dimensional feature X of the cell (v) The multi-dimensional low-dimensional representation H is calculated using residual connection through a nonlinear layer, group normalization, and a GELU activation function (v) ; S130. Set the cover set M, which consists of X (v) The index of the medium and low quality data, and the multi-dimensional fusion representation H is extracted for the cells not covered by the cover (f) ; S140. Multi-fractal low-dimensional representation H (v) Multi-fractal remodeling representation of the fractally characteristic cells in the uncovered obstacle And build remodeling loss L rec Multi-fractal fusion representation H using the fractal of the covered obstacle (f) Generate cross-fractal prediction representation And build prediction loss L pres ; S150. Optimizing the initial partition loss L p1 , H (v) and H (f) Obtaining the initial partition Q of cells using a cell partitioner (v) ; S160. Calculate the remodeling error, partition entropy, suspension, and confidence weight of the multi-dimensional representation of the cell; S170. Constructing a prediction loss L pres , a partition coordination loss L par and a partition entropy loss L paren , optimizing a partition loss L p2 training a model to convergence; S180. Partition Q of cells from the converging model (v) The final cell partition is calculated, and the accuracy of the cell cluster partition is obtained by combining the real labels of the cells in the dataset. In the S110, the cell multi-dimensional feature X (v) The multi-dimensional feature of the i-th cell is defined as follows: where n is the total number of cells, d v is the characteristic dimension; In S130, the X (v) After using PCA, the expression median deviation of each gene is calculated. If the number of genes in the vth dimension of the ith cell that exceed the expression median deviation threshold exceeds a set threshold, the dimension and the index value of the cell are recorded in the cover set M, which is defined as follows: If (i, v) e M, then Coated obstacles, whose data features are not available; The feature extraction multi-fractal fusion representation H of the uncoated obstacle cells is defined as follows: (f) , defined as follows: wherein, and are H (f) and H (v) is the low-dimensional representation of the corresponding i-th cell in the i-th group, F() maps to a high-dimensional tensor by performing an outer product operation and then using an attention mechanism to obtain the input components into an overall variable, and is composed of H (f) ; In the S140, the multi-dimensional reconstruction of the cell indicates is defined as follows: wherein D (v) The input is extracted features through two nonlinear layers, an average pooling layer, and an activation function ReLU. Remodeling loss L rec is defined as follows: where V is the total number of fractal dimensions, is the element-wise multiplication operation, γ (v) is the dimension weight, is the learnable parameter vector, and ∈ is a non-negative parameter. Cross fractal predictive representation is defined as follows: Predicted loss L pres is defined as follows: In the S150, the initial division loss L p1 is defined as follows: L p1 = L rec + λ1L pres Wherein, λ1 is a balance parameter; Objective: Minimize L p1 Cell initial division Q is obtained by a cell divider (v) The implementation of the cell divider is defined as follows: wherein K is the total number of cell classes, is the case of the ith cell belonging to each class in the vth fractal dimension, softmax() is a normalization activation function, linear() is a linear function, τ is a controllable hotness index, f φ is an inference network; In the S160, the remodeling error expressed by the i-th cell in the v-th fractal dimension is defined as follows: the partition entropy of the i-th cell in the v-th fractal dimension is defined as follows: wherein, is the kth column of the matrix of the vth fractal dimension, indicates the likelihood that the ith cell belongs to the class k. Suspension u of the i-th cell division in the v-th fractal dimension i,v Is defined as follows: Wherein, α is a weight coefficient, and norm() is a normalization operation; confidence weight for the i-th cell division in the v-th fractal dimension is defined as follows: Wherein, β is a heat scaling hyperparameter; In the S170, the coordination loss L is divided par is defined as follows: Wherein, || || 2 is an L2 norm; partitioning entropy loss L paren is defined as follows: Constructing the split loss L p2 is defined as follows: L p2 = L p1 + λ2L par + λ3L paren where λ2and λ3are balancing parameters, L is minimized by Adam p2 training the model to convergence; In the S180, the calculation method of the final cell partition is as follows: where, is the comprehensive partition of the i-th cell in all fractal dimensions, q i is the final class label of the i-th cell, which is is the class label corresponding to the maximum value in, and combining the real label of the cell in the dataset gives the accuracy of the cell class partition.

Images