Dimensionality reduction method for single-cell RNA sequence dataset

Abstract

The application discloses a dimension reduction processing method for single-cell RNA sequence data sets, and comprises the following steps: searching for k nearest neighbors (KNN) of each cell and counting RNN through the KNN; traversing the cells in descending order of RNN, and if the point is extracted, the KNN of the cell is removed until all cells are traversed, and then the anchor point cells obtained by sampling are output; constructing a KNN graph network of all anchor point cells, calculating the shortest path between all anchor point cells, and calculating their high-dimensional probability distribution; reducing all anchor point cells to a low-dimensional space and constantly updating the low-dimensional coordinates until the iteration condition is met and the iteration is terminated; for each non-anchor point cell that is not sampled, the mapping relationship between the low-dimensional coordinates before and after updating of d+1 anchor point cells is used to calculate the low-dimensional space coordinates of the non-anchor point cell; and integrating the low-dimensional coordinates of all anchor point cells and non-anchor point cells, and outputting the low-dimensional coordinates. The application greatly improves the dimension reduction efficiency and reduces the complexity of calculation.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of data processing, and particularly relates to a dimension reduction processing method for single-cell RNA sequence data sets. BACKGROUND

[0002] Dimension reduction is a machine learning method for data feature simplification and visualization in the field of bioinformatics. Single-cell RNA sequence (scRNA-seq) data analysis often faces tasks such as data cleaning, feature engineering and cell integration. There are often a large number of redundant or linearly correlated genes in scRNA-seq data sets, which not only affect the judgment of gene significance and the recognition of cell types, but also greatly increase the storage and computing cost. Dimension reduction aims to map high-dimensional scRNA-seq data to a low-dimensional space, eliminate redundant or highly correlated genes, and screen high-variability genes, thereby realizing the simplification of scRNA-seq data, so as to reduce the time and space complexity of subsequent calculations. In addition, dimension reduction can reduce the influence of invalid and erroneous data on subsequent processing and analysis, and improve the accuracy and robustness of the model. At the same time, dimension reduction (to 2 or 3-dimensional space) is also an important technical means for the visualization analysis of high-dimensional scRNA-seq data.

[0003] The existing dimension reduction methods can be roughly divided into linear and nonlinear methods. The classical linear methods are most widely used, such as PCA and LDA (Linear Discriminant Analysis), but this kind of method is difficult to separate different cell types in low-dimensional space for non-linear data distribution, such as island distribution, thus affecting the subsequent cell classification or clustering accuracy. Nonlinear methods include MDS (Multi-Dimensional Scaling), IsoMap (Isometric Feature Mapping), t-SNE (t-Distributed Stochastic Neighbor Embedding), Picasso and UMAP (Uniform Manifold Approximation and Projection). The goal of MDS is to minimize the distance difference of data in high and low dimensional space, but for the data of manifold distribution, the traditional Euclidean distance cannot well depict the nonlinear distribution of data, leading to the crowded phenomenon in low-dimensional space. IsoMap selects the path distance to optimize the traditional Euclidean distance, but still cannot effectively guarantee the separation of different cell types in low-dimensional space. t-SNE uses t-student distribution function to make the objects close in high-dimensional space closer in low-dimensional space, and the objects far away are farther away, which can better realize the separation of different cell types, but its high time complexity makes it difficult to perform dimension reduction task on large-scale scRNA-seq data set. Picasso uses autoencoder to make data embedded in any shape distribution, but its time efficiency is very low. UMAP improves the objective function of t-SNE to speed up the gradient convergence, but it is easy to ignore the global features of data distribution. SUMMARY

[0004] The purpose of the present application is to solve the problem of high computational complexity of traditional dimension reduction methods in processing scRNA-seq data set, greatly improve the dimension reduction efficiency of scRNA-seq data set, and better gather adjacent cells in low-dimensional space to separate different cell types from each other.

[0005] To solve the above technical problems, the technical scheme adopted by the present application is as follows:

[0006] A dimension reduction processing method for single cell RNA sequence data set, comprising the following steps:

[0007] S1, obtain a single cell RNA sequence data set, pre-process the data set to obtain all cells to be processed for dimension reduction, search for k nearest neighbors (KNN) of all cells, and count the number of times each cell is taken as a nearest neighbor (RNN) through the KNN;

[0008] S2, arrange all cells in descending order according to the size of RNN, and traverse all cells in turn, if the point is extracted, remove all KNN cells of the cell, until all cells are traversed, and output the anchor cells obtained by sampling;

[0009] S3, construct a KNN graph network of all anchor points, calculate the shortest path distance between all anchor cells in the KNN graph network to represent the similarity of the anchor cells in the manifold space, and calculate the high-dimensional probability distribution between the two according to the shortest path distance;

[0010] S4, reduce all anchor cells to a low-dimensional space, establish a low-dimensional probability distribution function of the anchor cells, and continuously update the low-dimensional coordinates of the anchor cells according to the high-dimensional probability distribution of the anchor cells until the iteration condition is met to terminate the iteration;

[0011] S5, for each non-anchor cell not sampled, first search for the nearest d+1 anchor cells, map the non-anchor cell to the low-dimensional space, and calculate the low-dimensional space coordinates of the non-anchor cell according to the mapping relationship between the low-dimensional coordinates before and after updating of the d+1 anchor cells in step 4;

[0012] S6, integrate the low-dimensional coordinates of all anchor cells and non-anchor cells, and output the low-dimensional coordinates of all cells, thereby realizing dimension reduction processing of the single cell RNA sequence data set.

[0013] Further, step 1 specifically includes:

[0014] S11, construct a KD index of all cells, and search for k nearest neighbors (KNN) of each cell based on the KD index;

[0015] S12, count the total number of times each cell is taken as a KNN object by other cells as its RNN value.

[0016] Further, the specific method of step 3 is:

[0017] S31, construct a KNN graph network of all anchor cells according to their KNN;

[0018] S32, in the KNN graph network, connect the two cells with the shortest distance between the disconnected subgraphs to construct a complete connected graph;

[0019] S33, using Floyd algorithm to calculate the shortest path distance between each pair of anchor cells in step 32 to represent the similarity of anchor cells in manifold space;

[0020] S34, calculating the high-dimensional probability distribution of anchor cells according to the shortest path distance:

[0021]

[0022] wherein, x i and x j respectively correspond to the coordinates of the i-th and j-th anchor cells in high-dimensional space, σ i is the bandwidth smoothing parameter of Gaussian distribution, and Φ(x i , x j ) represents the shortest path distance between the coordinates x i and x j , which represents the similarity of anchor cells in manifold space.

[0023] Further, step 4 specifically includes the following sub-steps:

[0024] S41, using PCA to reduce all anchor cells to d-dimensional low-dimensional space, establishing the low-dimensional probability distribution function of anchor cells in low-dimensional space, and initializing the d-dimensional coordinates of anchor cells:

[0025] q ij = (1 + s log (1 + ||y i -y j || 2 )) -1 ;

[0026] wherein, y i and y j respectively correspond to the coordinates of the i-th and j-th anchor cells in low-dimensional space, q ij represents the low-dimensional probability distribution of any anchor cell, and S is the aggregation parameter;

[0027] S42, using KL divergence as the loss function, using gradient descent method to calculate the gradient of low-dimensional coordinates of anchor cells combined with low-dimensional probability distribution and high-dimensional probability distribution, and constantly updating the low-dimensional coordinates of anchor cells:

[0028]

[0029]

[0030] wherein, is the gradient of the loss function KL divergence to y i , y i tis the updated low-dimensional coordinate of anchor cell i in the tth iteration, and a and b respectively correspond to learning rate and dynamic parameter;

[0031] S43, in the process of updating the low-dimensional coordinates of anchor cells, whether to end the iteration is determined by judging whether the loss function KL divergence is no longer reduced or the number of iterations reaches the target number T.

[0032] Further, step 5 specifically comprises:

[0033] S51, for each non-anchor cell, search the nearest d+1 anchor cells, and project them into the d-dimensional space by the PCA method;

[0034] S52, calculate the best scaling scale between the coordinates obtained by the PCA dimension reduction of the d+1 anchor cells and the coordinates obtained by step 4, and calculate the coordinates of the PCA coordinates of the anchor cells after scaling;

[0035]

[0036] X=X PCA *scale;

[0037] Wherein, d ij is the point pair distance calculated by the PCA coordinates of the ith anchor cell and the jth anchor cell, D ij is the point pair distance calculated by the coordinates obtained by step 4 of the ith cell and the jth cell, scale is the scaling scale, X PCA is the coordinate obtained by the PCA dimension reduction of the anchor cell, and X is the scaled PCA coordinate of the anchor cell;

[0038] S53, according to the coordinates obtained by step 4 and the coordinates obtained by step 52 of the anchor cell, the low-dimensional coordinates of the non-anchor cell are calculated, specifically:

[0039]

[0040] Wherein, y i , y j respectively correspond to the low-dimensional coordinates of the ith and jth anchor cells obtained by step 4, x i , x j respectively correspond to the low-dimensional coordinates of the ith and jth anchor cells obtained by step 52, and y is the low-dimensional coordinate of the non-anchor cell.

[0041] Further, the low-dimensional coordinates of all anchor cells and non-anchor cells are integrated, and the low-dimensional coordinates of all points are output according to the sorting of the original data.

[0042] Compared with the prior art, the beneficial effects of the present application are:

[0043] 1) The present application selects anchor cells in a sampling manner, and then performs manifold learning on the anchor cells to obtain the coordinates of the anchor cells in a low-dimensional space, and then embeds non-anchor cells into the low-dimensional space according to the relative relationship with the anchor cells, thereby greatly improving the dimension reduction efficiency of the data set, reducing the complexity of the calculation, solving the problem of high computational complexity when the traditional dimension reduction method is used for dimension reduction processing of the scRNA-seq data set; in addition, the shortest path distance is used instead of the traditional Euclidean distance, which can effectively process the manifold distribution structure of the cells; and by establishing a low-dimensional probability distribution function, the adjacent cells can be better aggregated in the low-dimensional space, solving the congestion problem of the cells in the low-dimensional space caused by the prior art, so that different cell types can be separated from each other;

[0044] 2) The method can quickly reduce the dimension of large-scale scRNA-seq data, and is applied to the visualization of biological single-cell data set, and the dimension-reduced result can be used as a preprocessing step for single-cell classification, facilitating better separation of different cell types in the subsequent process. BRIEF DESCRIPTION OF DRAWINGS

[0045] Figure 1 is the algorithm flowchart of the data dimension reduction processing method of the embodiment of the present application;

[0046] Figure 2 is the standard biological processing flowchart of the original scRNA-seq data set of the embodiment of the present application;

[0047] Figure 3 is the shortest path distance calculation flowchart of the embodiment of the present application;

[0048] Figure 4 is the fast embedding algorithm flowchart of the embodiment of the present application;

[0049] Figure 5 is the dimension reduction result comparison chart in the embodiment of the present application;

[0050] Figure 6 is the clustering accuracy comparison chart obtained according to the results of three dimension reduction methods in the embodiment of the present application;

[0051] Figure 7 is the running time comparison chart of the scRNA-seq data set in the embodiment of the present application. DETAILED DESCRIPTION

[0052] The technical solutions in the embodiments of the present application will be clearly and completely described below with the embodiments of the present application. Obviously, the described embodiments are only a part of the embodiments of the present application, but not all the embodiments of the present application. Based on the embodiments of the present application, all other embodiments obtained by a person of ordinary skill in the art without creative work belong to the protection scope of the present application.

[0053] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict.

[0054] In order to solve the problems of high computational complexity and crowded low-dimensional space when the existing dimension reduction method is used to process the dimension reduction task of the scRNA-seq data set, a single cell RNA sequence data set dimension reduction processing method is disclosed, comprising the following steps:

[0055] S1, obtaining a single cell RNA sequence data set, pre-processing the data set to obtain all cells to be dimensionally reduced, forming a cell set, searching for k nearest neighbors KNN of all cells in the cell set, and counting the number of times RNN that each cell is taken as a nearest neighbor by KNN; the step specifically comprises:

[0056] S11, pre-processing the data in the data set includes eliminating low-quality cells and excessively high gene counts due to double or multiple duplexes, and using a global scaling normalization method to normalize the gene expression measurement value of each cell; then, the mean and variance of the expression level of each gene in all cells are calculated, and the top several high-variation genes are selected; finally, PCA is adopted to screen out the cells to be dimensionally reduced and their principal components;

[0057] S12, constructing the KD index of all cells to be dimensionally reduced, and searching for k nearest neighbors KNN of each cell based on the KD index;

[0058] S13, counting the total number of times that each cell is taken as a KNN object by other cells as its RNN value.

[0059] S2, arranging all cells in descending order according to the size of RNN, and traversing all cells in turn, if the point is extracted, eliminating all KNN cells of the cell, until all cells are traversed, and outputting the anchor cells obtained by sampling; specifically,

[0060] S21, arranging all cells in descending order according to the RNN value of the cell;

[0061] S22, traverse the cells in the queue according to the arrangement order of step 21, if a cell is sampled, remove all KNN cells of the cell from the queue, and then continue to traverse until the sampling is completed, output the anchor cells obtained by sampling, and the cells in the queue (not sampled) are non-anchor cells.

[0062] S3, construct the KNN graph network of all anchor cells, calculate the shortest path distance between each pair of anchor cells in the KNN graph network to represent the similarity of the anchor cells in the manifold space, and calculate the high-dimensional probability distribution between them according to the shortest path distance; the specific method of this step is:

[0063] S31, construct the KNN graph network of all anchor cells according to the KNN of all anchor cells;

[0064] S32, in the KNN graph network, connect the two cells with the shortest distance between subgraphs for the disconnected subgraphs to construct a complete connected graph;

[0065] S33, calculate the shortest path distance between each pair of anchor cells in the complete connected graph of step 32 by using Floyd algorithm, which represents the similarity of the anchor cells in the manifold space; compared with the traditional Euclidean distance, the shortest path distance can better depict the similarity of the cells in the manifold space, and is more conducive to embedding the nonlinear manifold distribution into the low-dimensional space;

[0066] S34, calculate the high-dimensional probability distribution of the anchor cells according to the shortest path distance:

[0067]

[0068] In the formula, x i , x j respectively correspond to the coordinates of the i-th and j-th anchor cells in the high-dimensional space, σ i is the bandwidth smoothing parameter of the Gaussian distribution, Φ(x i , x j ) represents the shortest path distance between the coordinates x i and the coordinates x j , which represents the similarity of the anchor cells in the manifold space.

[0069] S4, reduce all anchor cells to a low-dimensional space, construct a low-dimensional probability distribution function of the anchor cells in the low-dimensional space, and update the low-dimensional coordinates of the anchor cells by combining the high-dimensional probability distribution of the anchor cells until the iteration condition is met to terminate the iteration; the specific method of this step is:

[0070] S41, reduce all anchor cells to a d-dimensional low-dimensional space by using PCA, establish a low-dimensional probability distribution function of the anchor cells in the low-dimensional space, and initialize the d-dimensional coordinates of the anchor cells:

[0071] q ij = (1 + s log (1 + ||y i -y j || 2 )) -1 ;

[0072] where y i , y j correspond to the coordinates of the i-th and j-th anchor cell in the low-dimensional space, q ij represents the probability distribution of any anchor cell in the low-dimensional space, and S is the clustering parameter;

[0073] S42, using the KL divergence as the loss function, the gradient of the low-dimensional coordinates of the anchor cells is calculated by using the gradient descent method combined with the low-dimensional probability distribution and the high-dimensional probability distribution, and the low-dimensional coordinates of the anchor cells are constantly updated:

[0074]

[0075]

[0076] where is the gradient of the loss function KL divergence with respect to y i , y i t is the updated low-dimensional coordinate of the anchor cell i in the t-th iteration, and a and b correspond to the learning rate and the dynamic parameter, respectively;

[0077] S43, in the process of updating the low-dimensional coordinates of the anchor cells, whether the loss function KL divergence is no longer reduced or the number of iterations reaches the target number T is judged to determine whether the iteration is ended.

[0078] S5, for each unsampled non-anchor cell, the d+1 nearest anchor cells are searched first, the non-anchor cell is mapped to the low-dimensional space, and the low-dimensional space coordinates of the non-anchor cell are calculated according to the mapping relationship between the low-dimensional coordinates before and after the update of the d+1 anchor cells in step 4; this step includes the following sub-steps:

[0079] S51, for each non-anchor cell, search for the d+1 nearest anchor cells, and project it to the d-dimensional space by the PCA method;

[0080] S52, calculate the best scaling scale between the coordinates obtained by PCA dimension reduction of the d+1 anchor cells and the coordinates obtained by step 4, and calculate the coordinates of the PCA coordinates of the anchor cells after scaling;

[0081]

[0082] X = X PCA * scale;

[0083] wherein, d ij is the point pair distance calculated by the PCA coordinates of the i th anchor cell and the j th anchor cell, D ij is the point pair distance calculated by the coordinates obtained in step 4, scale is a scaling scale, X PCA is the coordinate obtained by PCA dimension reduction of the anchor cell, and X is the scaled PCA coordinate of the anchor cell;

[0084] S53, calculate the low-dimensional coordinates of the non-anchor cell according to the coordinates obtained in step 4 and the coordinates obtained after scaling in step 52, specifically:

[0085]

[0086] wherein, y i , y j respectively correspond to the low-dimensional coordinates of the i th and j th anchor cell obtained in step 4, x i , x j respectively correspond to the low-dimensional coordinates of the i th and j th anchor cell obtained after scaling in step 52, and y is the low-dimensional coordinate of the non-anchor cell.

[0087] S6, integrate the low-dimensional coordinates of all anchor cells and non-anchor cells, and output the low-dimensional coordinates of all cells according to the sorting of the original data. After obtaining the low-dimensional coordinates of all cells, different cell types are identified according to the clustering of the cells in the low-dimensional space, and the heterogeneity between cells is revealed; and the differentiation level of the cells is analyzed according to the distribution of the cells and the positional relationship between them, combined with the cell differentiation tree.

[0088] The application will be further described below in combination with specific examples, but not as a limitation of the application.

[0089] The embodiment of the application discloses a data dimension reduction processing method, characterized by comprising the following steps:

[0090] There are four published scRNA-seq datasets with real cell type labels: WT_R1, WT_R2, NdpKO_R1, NdpKO_R2, which are from adult mouse retinas and normoxic and hypoxic cerebral cortex cells, containing 7551, 7695, 9939 and 7640 cells, respectively, and the expression levels of cells are tested on 27998 genes. The above datasets need to be dimensionally reduced to reveal the heterogeneity and differentiation level between cells, and further identify different cell types. In order to make different cell types have better separation effect in low-dimensional space, the aggregation parameter s of the low-dimensional probability distribution function is set to 30, and of course other values can be selected in other embodiments, which is set according to actual needs. When calculating the gradient, the stochastic gradient descent method is adopted, and 30% of the anchor cells are randomly selected each time to calculate and update the gradient to improve the calculation efficiency of gradient descent. Among them, the iteration number, learning rate and dynamic parameter are set to T=1000, α=n / 20 (n is the total number of cells) and β=1, respectively.

[0091] The dimension reduction process of the embodiment is shown in Figure 1 , specifically, the multi-order neighborhood uniform sampling method is adopted to reduce the number of cells participating in dimension reduction calculation, and the low-dimensional probability distribution function and the stochastic gradient descent method are optimized to speed up the convergence of the loss function; finally, the fast embedding method is adopted to efficiently map non-anchor cells to low-dimensional space, thereby realizing fast dimension reduction of large-scale scRNA-seq datasets. The specific steps of the embodiment are as follows:

[0092] 1) The standard biological processing flow for the original scRNA-seq dataset is shown in Figure 2 . scRNA-seq stores data in single molecule tag (UMI) matrix format, each row represents a cell, and each column represents a gene. First, the UMI matrix is subjected to quality control processing, and cells with low quality and genes with excessively high gene counts due to double or multiple duplexes are removed. Second, the global scaling normalization method is used to normalize the gene expression measurement value of each cell; then, the mean and variance of the expression level of each gene in all cells are calculated, the top 2000 high-variation genes are selected, and finally the PCA is adopted to select the top 50 principal components of the cells;

[0093] 2) Obtain the cell set composed of the screened cells, search for the 30 cells (not limited to 30) closest to each cell in the cell set as its KNN value, and count the corresponding RNN value. All cells are arranged in descending order of RNN, and are sampled one by one. After each sampling, the KNN cells are removed from the queue, and the sampling is completed until the anchor cells are output;

[0094] 3) The flow of calculating the shortest path distance of all anchor cells is shown in Figure 3As shown, firstly, the KNN graph network is constructed by KNN search, the nearest cell connection between any two disconnected subgraphs is connected to form a globally connected graph network, and finally the Floyd algorithm is used to find the shortest path between any two cells in the globally connected graph, and the high-dimensional probability distribution between the two is calculated according to the shortest path distance;

[0095] 4) Using KL divergence as the loss function, the low-dimensional probability distribution function of the anchor cell in the low-dimensional space is constructed, combined with the anchor high-dimensional probability distribution, and the gradient descent method is used to calculate the low-dimensional coordinates of the anchor cell; the logarithmic function is used to model the low-dimensional space probability distribution, which can better gather adjacent cells in the low-dimensional space than exponential function and t distribution function, and different cell types are separated from each other;

[0096] 5) In order to quickly embed the non-anchor cell into the low-dimensional space determined by the anchor cell, the PCA dimension reduction is performed on the non-anchor cell and the adjacent d+1 anchor cells, and then scaled to the same scale as the low-dimensional space of the anchor cell, and finally the two are matched to calculate the coordinates of the non-anchor cell in the low-dimensional space, and the specific algorithm process is as shown in Figure 4 ;

[0097] 6) Finally, the low-dimensional coordinates of the anchor cell and the non-anchor cell are integrated, and the dimension reduction results of the mouse retina and cerebral cortex cells are output as shown in Figure 5 ; Figure 5 It can be found from Figure 6 that the embodiment can better preserve the data structure of each cluster than the traditional dimension reduction method (such as Picasso and t-SNE), and can also maintain the separation of different category clusters, and the embodiment also solves the problem of cell congestion in the low-dimensional space compared with the traditional dimension reduction method. The clustering accuracy obtained by inputting the results of the three dimension reduction methods into the K-means clustering method is as shown in Figure 6 ; From the clustering results shown in Figure 7 , the clustering accuracy of the embodiment is higher, which is more conducive to subsequent clustering or classification operations. In addition, the method of the embodiment selects anchor cells in a sampling manner, and then performs manifold learning on the anchor cells, and from the running time shown in , the present application has obvious time efficiency advantage on other 12 scRNA-seq data sets than other methods (such as t-SNE and UMAP), and has lower computational complexity.

[0098] The above is only the preferred embodiment of the present application, and does not limit the implementation and protection scope of the present application. For those skilled in the art, it should be realized that any equivalent replacement and obvious changes made by applying the content of the present application should be included in the protection scope of the present application.

Claims

1. A method for dimension reduction processing of single-cell RNA sequence data sets, characterized by, The method comprises the following steps: S1, obtaining a single cell RNA sequence data set, preprocessing the data set to obtain all cells to be processed for dimension reduction, searching for k nearest neighbors (KNN) of all cells, and counting the number of times each cell is taken as a nearest neighbor (RNN) through the KNN; S2, arranging all cells in descending order according to the size of RNN, and traversing all cells in turn, if the point is extracted, removing all KNN cells of the cell, until all cells are traversed, and outputting anchor cells obtained by sampling; S3, constructing a KNN graph network of all anchor cells, calculating the shortest path distance between all anchor cells according to the KNN graph network to represent the similarity of the anchor cells in the manifold space, and calculating the high-dimensional probability distribution of the anchor cells according to the shortest path distance; S4, dimensionally reducing all anchor cells to a low-dimensional space, establishing a low-dimensional probability distribution function of the anchor cells and continuously updating the low-dimensional coordinates of the anchor cells according to the high-dimensional probability distribution of the anchor cells until the iteration condition is met to terminate the iteration; specifically comprising the following sub-steps: S41, using PCA to dimensionally reduce all anchor cells to a d-dimensional low-dimensional space, establishing a low-dimensional probability distribution function of the anchor cells in the low-dimensional space, and initializing the d-dimensional coordinates of the anchor cells: ； wherein, respectively correspond to the coordinates of the first , the second anchor cell in the low-dimensional space, represents the probability distribution of any anchor cell in the low-dimensional space, S is a concentration parameter; S42, using KL divergence as a loss function, using gradient descent method to calculate the gradient of the low-dimensional coordinates of the anchor cells combined with the low-dimensional probability distribution and the high-dimensional probability distribution, and continuously updating the low-dimensional coordinates of the anchor cells: ； ； wherein, is the gradient of the loss function KL divergence with respect to , is the anchor cell the low-dimensional coordinates updated in the t th iteration, α and β respectively correspond to the learning rate, the momentum parameter; S43, in the process of updating the low-dimensional coordinates of the anchor cells, whether the loss function KL divergence is no longer reduced or the iteration number reaches the target number T is judged to determine whether the iteration is ended; S5, for each non-anchor cell not sampled, searching for the nearest d+1 anchor cells, mapping the non-anchor cell to the low-dimensional space, and calculating the low-dimensional space coordinates of the non-anchor cell according to the mapping relationship between the d+1 anchor cells before and after updating the low-dimensional coordinates in step 4; S6, integrating the low-dimensional coordinates of all anchor cells and non-anchor cells, outputting the low-dimensional coordinates of all cells, thereby realizing the dimension reduction processing of the single cell RNA sequence data set.

2. The method of claim 1, wherein the method is performed by a computer system. Step 1 specifically comprises: S11, constructing a KD index of all cells, searching for k nearest neighbors (KNN) of each cell based on the KD index; S12, counting the total number of times each cell is taken as a KNN object by other cells as its RNN value. 3.The method of claim 1, wherein, The specific method of step 3 is: S31, constructing a KNN graph network of all anchor cells according to the KNN; S32, in the KNN graph network, connecting the two cells with the shortest distance between the disconnected subgraphs to construct a complete connected graph; S33, using Floyd algorithm to calculate the shortest path distance between all anchor cells in the complete connected graph in step 32 to represent the similarity of the anchor cells in the manifold space; S34, calculating the high-dimensional probability distribution of the anchor cells according to the shortest path distance: ； In the formula, respectively correspond to the coordinates of the first , the second anchor cell in high-dimensional space, is the bandwidth smoothing parameter of the Gaussian distribution, represents the coordinates and the coordinates The shortest path distance between them characterizes the similarity of the anchor cells in the manifold space. 4.The method of claim 1, wherein, Step 5 specifically comprises: S51, for each non-anchor cell, search the d+1 anchor cells closest to it, and project it to d-dimensional space using PCA method; S52, calculate the best scaling between the coordinates of the d+1 anchor cells obtained by PCA dimension reduction and the coordinates obtained by step 4, and calculate the scaled coordinates of the PCA coordinates of the anchor cells; ； in, d ij It is the first i The anchor cell and the first j Point-to-point distances calculated from the PCA coordinates of each anchor cell. D ij It is the first i The cell and the first j The point-to-point distance is calculated from the coordinates of each cell obtained in step 4. scale It's a scaling factor. X PCA These are the coordinates obtained by PCA dimensionality reduction of the anchor cells. X These are the PCA coordinates of the anchor cells after scaling. S53, calculate the low-dimensional coordinates of this non-anchor cell according to the coordinates of the anchor cells obtained in step 4 and the scaled coordinates obtained in step 52, specifically: ； wherein, y i 、y j respectively correspond to the low-dimensional coordinates of the first i j x i 、x j respectively correspond to the low-dimensional coordinates of the first i j y is the low-dimensional coordinate of the non-anchor cell.​​​​ 5. The method of claim 1, wherein, Integrate the low-dimensional coordinates of all anchor cells and non-anchor cells, and output the low-dimensional coordinates of all cells according to the original data sorting.

Images