Method for dividing cell types based on single-cell multi-omics data

Abstract

The invention relates to a method for dividing cell types based on single-cell multi-omics data, and belongs to the technical field of bioinformatics. The problems of high noise and high sparsity formed by a common technology are solved. Comprising the following steps: step 1, constructing a deep neural network model, step 2, obtaining a single-cell multi-omics comprehensive data set, step 3, preprocessing the single-cell multi-omics comprehensive data set, step 4, training the deep neural network model based on the preprocessed single-cell multi-omics comprehensive data set, and step 5, training the deep neural network model based on the preprocessed single-cell multi-omics comprehensive data set. And performing cell type division on to-be-detected data based on the trained deep neural network model. According to the method, an scDRMAE model is provided, dependency relationships of different omics are captured through two parallel MAE modules, factors which are usually ignored are considered in cell clustering, interpolation of different omics data is realized by means of a mask prediction mechanism in the cell clustering, and finally, the omics data are effectively fused through an attention mechanism, so that the accuracy of cell clustering is improved. And the accuracy of model cell clustering is improved.

Description

Technical Field

[0001] The present invention relates to a method for cell type classification based on single-cell multi-omics data, and belongs to the technical field of bioinformatics. Background Art

[0002] Advances in single-cell RNA sequencing (scRNA-seq) technology have enabled biologists to measure genome-wide expression profile data at the single-cell level, thereby effectively dissecting heterogeneous cell populations in complex samples that are not accessible through traditional bulk sequencing methods. The rapid development of single-cell technology has provided a unique perspective for understanding cellular heterogeneity and dynamic changes in complex biological systems. In recent years, the emergence of emerging multimodal sequencing technologies such as CITE-seq and REAP-seq, which combine transcriptome sequencing with single-cell surface protein determination, has enabled researchers to simultaneously study gene transcription levels and protein expression levels at the single-cell level. Specifically, CITE-seq uses existing single-cell sequencing technology to quantitatively determine the abundance of cell surface proteins by counting antibody-derived tags ADT; while REAP-seq combines DNA barcode antibodies with scRNA-seq methods to measure the expression levels of genes and cell surface proteins. Both methods have high sensitivity and can detect low-abundance protein expression, providing new possibilities for high-throughput single-cell analysis. In addition, the introduction of the single-cell chromatin accessibility sequencing technology scATAC-seq allows researchers to explore the accessibility of cellular chromatin, thereby revealing cell-type-specific gene expression regulation mechanisms. Technologies such as SNARE-seq and SMAGE-seq, which combine chromatin accessibility with gene expression analysis and allow simultaneous analysis of RNA expression and chromatin accessibility at the single-cell level, have been proposed to provide a new perspective for analyzing the complex biological processes of single cells. Recently, the three-peak analysis technology TEA-seq was proposed, which integrates transcriptome sequencing, single-cell surface protein determination, and chromatin accessibility sequencing technology, and is able to sequence and analyze three types of omics data simultaneously. The data generated by these technologies provide biologists with a more comprehensive characterization of cellular systems from a single-cell perspective, creating new opportunities for comprehensive analysis of cells. [4] .

[0003] Cluster analysis is an important basic step in most single-cell studies. It can identify cell subpopulations unsupervisedly and provide key support for downstream cell differential expression analysis and complex disease research. [5]. Over the years, methods for analyzing single-cell RNA sequencing (scRNA-seq) data have been continuously enriched and developed. However, due to the limited amount of information in a single omics data, especially when there are many cell subtypes, clustering methods developed based on scRNA-seq often fail to achieve ideal results. Therefore, further exploration of strategies and methods for integrating multi-omics data is expected to improve the accuracy and effectiveness of single-cell clustering analysis.

[0004] In multi-omics data analysis, the biological information provided by different omics data is complementary and can characterize cells from multiple perspectives. By integrating multi-omics data, the inherent information limitations of single omics data can be overcome, thereby achieving more accurate cell clustering. However, we note that most of the existing multi-omics clustering methods achieve clustering analysis through differences in expression patterns of different omics data, and usually do not fully consider the dependencies between different omics internal features. However, in different types of cells, the dependencies between different omics features often show significant differences, and these difference information may be helpful for cell clustering. In addition, scRNA-seq and scATAC-seq data are often accompanied by high noise and high sparsity, and dropout events occur, resulting in a large number of false zeros in the data.

[0005] The purpose of the present invention is to solve the problem of high noise and high sparsity formed by common technologies. A brief summary of the present invention is given below to provide a basic understanding of certain aspects of the present invention. It should be understood that this summary is not an exhaustive summary of the present invention. It is not intended to determine the key or important parts of the present invention, nor is it intended to limit the scope of the present invention. Summary of the invention

[0006] The technical solution of the present invention:

[0007] A method for cell type classification based on single-cell multi-omics data, characterized by comprising the following steps:

[0008] Step 1: Build a deep neural network model;

[0009] Step 2: Obtain a comprehensive single-cell multi-omics dataset;

[0010] Step 3: Preprocessing the single-cell multi-omics dataset;

[0011] Step 4: Train the deep neural network model based on the preprocessed single-cell multi-omics dataset;

[0012] Step 5: Classify the cell types of the test data based on the trained deep neural network model.

[0013] Preferably: the deep neural network model consists of an encoder, a mask decoder, a decoder and an attention fusion module;

[0014] The encoder, mask decoder, and decoder form a masked autoencoder;

[0015] Different omics encoders are used to extract important features of different omics data and the dependencies between features;

[0016] The mask decoder is used to decode the mask to predict the dropout event;

[0017] The decoder is used to reconstruct and interpolate dropout events.

[0018] Preferably: the single-cell multi-omics comprehensive dataset includes a multi-omics dataset consisting of scRNA-seq data and scATAC-seq data and a multi-omics dataset consisting of scRNA-seq data and scADT-seq data.

[0019] Preferably, the preprocessing of the single-cell multi-omics dataset specifically includes:

[0020] Step 1: For scRNA-seq data, we filtered out genes expressed in less than 1% of cells, screened 3,000 highly variable genes, and performed log normalization and standardization;

[0021] Step 2: For scATAC-seq data, remove the peaks present in less than 1% of cells, select 3000 highly variable peaks and perform binarization, and finally apply TF-IDF for feature transformation;

[0022] Step 3: Repeat step 2 to perform the same log normalization and standardization on the scADT-seq data.

[0023] Preferably, the training steps for the deep neural network model include:

[0024] The training steps for a deep neural network model include:

[0025] Step 1: Generate mask data and randomly shuffle some features in the original data matrix;

[0026] Step 2: Generate a mask matrix M whose items all obey the Bernoulli distribution RNA-seq , which is expressed as follows:

[0027] M RNA-seq ~Bernoulli(P) (1)

[0028] In formula (1), P represents the probability of dropout events occurring in different genes in different cells;

[0029] Therefore, by analyzing the raw data X of scRNA-seq RNA-seq , disrupt data and the mask matrix M RNA-seq Combine to get the masked data matrix x RNA-seq , which is expressed as follows:

[0030]

[0031] Similarly, the mask data x of other omics others , which is expressed as follows:

[0032]

[0033] In formula (3), M others represents the mask matrix of other omics, and others represents others;

[0034] Step 3: Mask decoder and Low-dimensional representation Z for learning from different omics RNA-seq and Z others The mask matrix corresponding to different omics data is predicted in , which is expressed as follows:

[0035]

[0036] In formula (4), and Represents the encoders of different omics, M' RNA-seq and M' others is the mask matrix of different omics predicted by the mask decoder;

[0037] By using binary cross entropy loss as the mask prediction loss L M , which enables the model to learn dropout events in omics data, as shown below:

[0038]

[0039] In formulas (6), (7), (8) and They represent the mask prediction losses of different omics, n represents the number of cells, g and p represent the feature dimensions of different omics data, and represents the real mask matrix, and represents the mask matrix predicted by the mask decoder;

[0040] The predicted mask matrices and low-dimensional representations of different omics are combined and passed into the decoder for reconstruction of different omics data, as shown below:

[0041]

[0042] In formulas (8) and (9), X' RNA-seq and X' others Represent the reconstruction data of different omics data, Z RNA-seq and Z others Represents the feature representation of scRNA-seq data after dimensionality reduction by the encoder;

[0043] The reconstruction loss L is calculated by calculating the mean square error between the reconstructed data and the original data of different omics R , while assigning different weights to damaged and undamaged genes, as follows:

[0044]

[0045] and Represent the reconstruction loss of different omics data, and Represent the reconstruction weights of perturbed and unperturbed features of scRNA-seq data and other omics data, respectively;

[0046] Step 4: Concatenate the low-dimensional representations of different omics, as shown below:

[0047] Z=[Z RNA-seq ,Z others ] (13)

[0048] The feature transformation of Z is performed through the attention mechanism, so that Z can pay more attention to the features that are important to the clustering results;

[0049] Using the same strategy as transformer, Z is mapped to three different feature subspaces:

[0050] K=ZW1; Q=ZW2; V=ZW3 (14)

[0051] In formula (14), W1, W2 and W3 represent three weight matrices used to map changes, and K, Q and V represent three feature codes after Z is linearly transformed;

[0052] Use K and Q to calculate the global relationship matrix W:

[0053]

[0054] The importance of different features is obtained through this formula, and softmax is used to map all values ​​​​to between 0 and 1

[0055] Where d represents the feature dimension of Z;

[0056] With the help of W, V is enhanced to obtain the enhanced features.

[0057]

[0058] b is the deviation;

[0059] By integrating the scRNA-seq data feature Z r Again and Fusion to obtain a more informative feature representation It is expressed as follows:

[0060]

[0061] The KL loss is introduced as a constraint, and the t distribution is used to describe the similarity between cells i and j:

[0062]

[0063] In formula (18) and represents the encoding representation of cells i and j, q ij represents the soft assignment, which is used to represent the pairwise similarity between cells i and j;

[0064] p ij Based on q ij The target distribution is constructed by ij After normalization, the calculation process is as follows:

[0065]

[0066] By finding the KL divergence of the two distributions, the features are constrained to improve the clustering effect:

[0067]

[0068] P represents the target distribution, Q represents the fused feature distribution, and our goal is to make Q close to P. This formula is used to calculate the divergence of the two distributions

[0069] Step 5: After obtaining the single-cell multi-omics fusion data, cluster analysis is performed on the single-cell multi-omics fusion data using the Kmeans algorithm.

[0070] The present invention has the following beneficial effects:

[0071] The present invention proposes a scDRMAE model, which captures the dependencies of different omics internal features through two parallel MAE modules, takes this usually ignored factor into account in cell clustering, and uses the mask prediction mechanism therein to achieve interpolation of different omics data. Finally, the attention mechanism is used to effectively fuse the various omics data, greatly improving the accuracy of the model cell clustering.

[0072] The scDRMAE model of the present invention can achieve excellent clustering performance on various types of multi-omics data sets. By comparing with other latest cell clustering models, the scDRMAE model shows better performance in multiple evaluation indicators;

[0073] The scDRMAE model constructed by the present invention also has strong noise resistance. In the face of dropout noise of different intensities, scDRMAE can still maintain better clustering performance compared with other methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0074] Figure 1 is a flow chart of a method for cell type classification based on single-cell multi-omics data;

[0075] Figure 2 It is the scDRMAE model diagram;

[0076] Figure 3 This is a comparison chart of clustering performance between scDRMAE and other methods on scRNA-seq and scATAC-seq datasets;

[0077] Figure 4 This is a comparison chart of clustering performance between scDRMAE and other methods on scRNA-seq and scADT-seq datasets;

[0078] Figure 5 This is the distribution map of the fusion features of scDRMAE on four data sets visualized by UMAP;

[0079] Figure 6 It is a comparison chart of clustering indicators of scDRMAE and other methods under different dropout rates;

[0080] Figure 7 It is a box plot of scDRMAE and other methods’ clustering indicators on the dataset;

[0081] Figure 8 It is the clustering comparison and visualization diagram of scDRMAE data before and after interpolation;

[0082] Fig. 9 This is a differential analysis diagram of cell clusters divided by scDRMAE. DETAILED DESCRIPTION

[0083] In order to make the purpose, technical scheme and advantages of the present invention clearer, the present invention is described below by the specific embodiments shown in the accompanying drawings. However, it should be understood that these descriptions are only exemplary and are not intended to limit the scope of the present invention. In addition, in the following description, the description of well-known structures and technologies is omitted to avoid unnecessary confusion of the concept of the present invention.

[0084] Specific implementation method 1: Combination Figure 1-Figure 9 This embodiment describes a method for cell type classification based on single-cell multi-omics data, including the following steps:

[0085] Step 1: Build a deep neural network model;

[0086] Step 2: Obtain a comprehensive single-cell multi-omics dataset;

[0087] Step 3: Preprocessing the single-cell multi-omics dataset;

[0088] Step 4: Train the deep neural network model based on the preprocessed single-cell multi-omics dataset;

[0089] Step 5: Classify the cell types of the test data based on the trained deep neural network model.

[0090] Specific implementation method 2: Combination Figure 1-Figure 9 This embodiment is described. This embodiment is a method for cell type classification based on single-cell multi-omics data. The deep neural network model consists of an encoder, a mask decoder, a decoder and an attention fusion module.

[0091] The encoder, mask decoder, and decoder form a masked autoencoder;

[0092] Different omics encoders are used to extract important features of different omics data and the dependencies between features;

[0093] The mask decoder is used to decode the mask to predict the dropout event;

[0094] The decoder is used to reconstruct and interpolate dropout events;

[0095] The attention fusion module consists of an attention mechanism and a residual-like network. The attention mechanism is used to dynamically adjust the weights of different omics data. The residual-like network is designed to prevent information loss during model training. The low-dimensional representation of scRNA-seq, a more important omics data, is reintegrated into the fused data after self-attention processing, thereby avoiding the degradation of cell clustering performance caused by information loss.

[0096] Specific implementation method three: Combination Figure 1-Figure 9 The present embodiment is described. The present embodiment is a method for cell type classification based on single-cell multi-omics data. The single-cell multi-omics comprehensive data set includes a multi-omics data set consisting of scRNA-seq data and scATAC-seq data and a multi-omics data set consisting of scRNA-seq data and scADT-seq data.

[0097] The human cell line mixture'GSE126074', mouse brain'GSE140203', bone marrow mononuclear cell dataset BMMC'GSE194122' and Ma-2020 dataset were downloaded from the GEO database, of which Ma-2020 contains 4 batches of data;

[0098] BMNC, InHouse'GSE148665', Stephenson, PBMC10k were obtained from the GEO database and previous studies. We divided the Stephenson dataset into two datasets, Ncl and Cambridge, according to different sequencing institutions.

[0099] Specific implementation method four: Combination Figure 1-Figure 9 This embodiment is described. This embodiment is a method for cell type classification based on single-cell multi-omics data. The steps of preprocessing the single-cell multi-omics data set include:

[0100] For scRNA-seq data, genes expressed in less than 1% of cells were filtered out, 3000 highly variable genes were screened and log normalized and standardized;

[0101] For scATAC-seq data, peaks present in less than 1% of cells were removed, 3000 highly variable peaks were selected and binarized, and finally TF-IDF was applied for feature transformation. scADT-seq data were also log-normalized and standardized.

[0102] Specific implementation method five: Combination Figure 1-Figure 9 This embodiment is described. In this embodiment, a method for cell type classification based on single-cell multi-omics data includes the following steps for training a deep neural network model:

[0103] Step 1: Generate mask data. In order to enable the model to better learn the dependencies between features, we randomly shuffle some features in the original data matrix, thereby introducing randomness in feature expression in different omics data.

[0104] Step 2: Secondly, according to existing research, the probability of each gene dropout event occurring is P = (p1, p2, p 3,…, p n ) usually obeys the Bernoulli distribution. In order to introduce dropout events into the input data and enhance the model's ability to learn dependencies between features, we generate a mask matrix M whose items obey the Bernoulli distribution. RNA-seq , which is expressed as follows:

[0105] M RNA-seq ~Bernoulli (P)

[0106] Where P represents the probability of dropout events occurring in different genes in different cells. RNA-seq , disrupt data And the mask matrix M RNA-seq Combine to get the final masked data matrix x RNA-seq , which is expressed as follows:

[0107]

[0108] Similarly, the mask data x of other omics others , which is expressed as follows:

[0109]

[0110] Step 3: Train the masked autoencoder. The masked autoencoder consists of two sets of parallel encoders, a masked decoder, and a decoder. The encoder is used to reduce the dimension of the masked data of different groups and project it into a low-dimensional space, which is expressed as follows:

[0111]

[0112] in and Represents the encoders of different omics, which are used to convert x RNA-seq and x others Projected to a low-dimensional space. In the mask decoder and Low-dimensional representation Z for learning from different omics RNA-seq and Z others The mask matrix corresponding to different omics data is predicted in , which is expressed as follows:

[0113]

[0114] Where M' RNA-seq and M' others is the mask matrix of different omics predicted by the mask decoder. By using binary cross entropy loss as the mask prediction loss LM , so that the model can learn the dropout events in the omics data. It is expressed as follows:

[0115]

[0116] and They represent the mask prediction losses of different omics, where n represents the number of cells, g and p represent the feature dimensions of different omics data, and represents the real mask matrix, and Represents the mask matrix predicted by the mask decoder. We merge the predicted mask matrices of different omics and low-dimensional representations into the decoder for reconstruction of different omics data. It is represented as follows:

[0117]

[0118]

[0119] where X' RNA-seq and X' others Represent the reconstructed data of different omics data. We calculate the mean square error between the reconstructed data of different omics and the original data as the reconstruction loss L R , while assigning different weights to damaged and undamaged genes, as follows:

[0120]

[0121] and Represent the reconstruction loss of different omics data, and Represent the reconstruction weights of perturbed and unperturbed features of scRNA-seq data and other omics data, respectively.

[0122] Step 4: Multi-omics data fusion. In order to further fuse the low-dimensional representations of different omics, we first concatenate the low-dimensional representations of different omics, as shown below:

[0123] Z=[Z RNA-seq ,Z others ]

[0124] Then, we transform the features of Z through the attention mechanism so that Z can pay more attention to the features that are important to the clustering results. First, we will use the same strategy as the transformer to map Z into three different feature subspaces:

[0125] K=ZW1;Q=ZW2;V=ZW3

[0126] Among them, W1, W2 and W3 represent three weight matrices used to map changes, and K, Q and V represent three feature codes after Z is linearly transformed. Then K and Q are used to calculate the global relationship matrix W:

[0127]

[0128] Where d represents the feature dimension of Z. Next, V is enhanced with the help of W to obtain the enhanced feature

[0129]

[0130] b is the deviation.

[0131] According to previous experience, scRNA-seq data can often provide richer information in the process of cell clustering. In order to prevent the degradation problem during model training, which leads to the loss of scRNA-seq information that is more important for clustering, we are inspired by ResNet and use the scRNA-seq data feature Z r Again and Fusion to obtain a more informative feature representation It is expressed as follows:

[0132]

[0133] In order to further optimize the fusion features In order to make information exchange between different omics MAE frameworks, we introduced KL loss as a constraint, which can make similar cells more compact and separate different types of cells from each other. Referring to previous studies, we use t distribution to describe the similarity between cell i and cell j:

[0134]

[0135] in and represents the encoding representation of cells i and j, q ij represents the soft assignment, which is used to represent the pairwise similarity between cells i and j. ij Based on q ij The target distribution constructed is q ij This construction is obtained by normalization. It aims to enhance the affinity between cells with higher similarity and reduce the affinity between cells with lower similarity. The calculation process is as follows:

[0136]

[0137] After obtaining the two distributions, we constrain the features by finding the KL divergence of the two distributions to improve the clustering effect:

[0138]

[0139] Step 5: After obtaining the single-cell multi-omics fusion data, we performed cluster analysis on the single-cell multi-omics fusion data using the Kmeans algorithm.

[0140] Example

[0141] scDRMAE was compared with the three existing multi-omics clustering methods of Moclust, scMCs, and scEMC, and the three single-omics clustering methods of Kmeans, scMAE, and scRISE. The cell clustering effects of these models were evaluated using the ARI, NMI, and AMI indicators. We conducted experiments on two types of datasets, one is a combination of scRNA-seq data and scATAC-seq data, and the other is a combination of scRNA-seq data and scADT-seq data. The first type of dataset contains 10 datasets, and the human cell line mixture'GSE126074', mouse brain'GSE140203', bone marrow mononuclear cell dataset BMMC'GSE194122' and Ma-2020 dataset were downloaded from the GEO database, of which Ma-2020 contains 4 batches of data. Secondly, for the combined datasets of scRNA-seq and scADT-seq, including BMNC, InHouse'GSE148665', Stephenson, PBMC10k, we divided the Stephenson dataset into two datasets, Ncl and Cambridge, according to different sequencing institutions.

[0142] Tables 1, 2, and 3 show the performance of the seven models on the three evaluation indicators in the first type of data set. Figure 3 The bar graphs of clustering performance evaluation indicators of these methods on four data sets are shown, including AMI, NMI and ARI. The results show that in the first type of data set, scDRMAE outperforms other compared methods in all three clustering indicators. In addition, we found that the performance of most multi-omics clustering methods is better than that of single-omics clustering methods. This result further verifies the effectiveness of using multi-omics data for cell clustering. This is attributed to the fact that these multi-omics data provide richer cell heterogeneity information than single-omics data, which helps to improve the performance of cell clustering. Therefore, it is very necessary to integrate multi-omics data for analysis.

[0143] In order to evaluate the generalization ability of the scDRMAE model, we conducted experiments on the second type of multi-omics dataset. Tables 4, 5, and 6 show the performance of the seven models on the three evaluation indicators in the second type of dataset. Figure 4 The bar graphs of clustering performance evaluation indicators of these methods on four datasets are shown, including AMI, NMI, and ARI. On the second category of datasets, except for the PBMC10k dataset, scDRMAE showed the best performance on the remaining four datasets, and won the second place on the PBMC10k dataset. In addition, Ncl and Cambridge are two multi-batch datasets, both covering peripheral blood mononuclear cell samples of 12 individuals, containing 48 cell subtypes and 45 cell subtypes, respectively, showing the advantages of the scDRMAE method in processing multi-batch datasets.

[0144] In order to intuitively demonstrate the clustering effect of scDRMAE, we visualized the fusion features of different data sets, such as Figure 5 shown. Figure 5 a is the UMAP graph of the InHouse dataset, which consists of 6 types of cells; Figure 5 b is the UMAP graph of the human cell line mixture data set, which consists of 4 types of cells. Figure 5 c and Figure 5 d is the UMAP graph of the mouse brain dataset and the MA-53 dataset, which contain 19 and 22 cell types, respectively. Overall, the separation effect of different major cell types in the four datasets is quite obvious, especially in the human cell line mixture and InHouse datasets with fewer cell subtypes. It is foreseeable that the separation effect of the mouse brain and MA-53 datasets is slightly worse than that of the first two datasets. This is because the mouse brain and MA-53 datasets contain more cell subtypes, and the expression patterns of features between different subtypes are often very similar, resulting in blurred boundaries between different cell subpopulations and unclear separation effects. Overall, the low-dimensional representation generated by scDRMAE can well distinguish different types of cell populations, further proving the superiority of scDRMAE clustering effect.

[0145] Table 1. ARI comparison of different methods on the first type of dataset

[0146]

[0147]

[0148] Table 2. AMI comparison of different methods on the first type of dataset

[0149]

[0150] Table 3. NMI comparison of different methods on the first type of data set

[0151]

[0152] Table 4. ARI comparison of different methods on the second type of dataset

[0153]

[0154] Table 5. NMI comparison of different methods on the second type of dataset

[0155]

[0156]

[0157] Table 6. AMI comparison of different methods on the second type of dataset

[0158]

[0159] To demonstrate the noise immunity of scDRMAE and its ability to learn dependencies between features, we simulate dropout events with different probabilities by multiplying different test sets with 0-1 mask matrices with different probabilities. Figure 6 The clustering performance of different methods under different dropout rates is demonstrated.

[0160] By comparing the clustering indicators of different methods under different dropout probabilities, it can be observed that with the increase of dropout probability, the clustering performance of all methods has decreased, among which scEMC has the largest fluctuation. It is worth noting that on the cell_line dataset, when the dropout rate rises to 90%, the ARI of scDRMAE can still reach 0.56, while the ARI, AMI and NMI of other methods such as scEMC are almost zero. Figure 7 Box plots of various clustering indicators using different methods are drawn. Figure 7a and 7b, we can see that the fluctuations of scDRMAE, scMCs, and Mocluster are small, but the clustering effect of scDRMAE is always better than these two methods at different dropout rates. The excellent performance of scDRMAE is attributed to the unique mask prediction mechanism of scDRMAE, which enables it to learn the dependencies between different features and effectively identify the occurrence of dropout events. Therefore, even at extremely high dropout rates, scDRMAE can still maintain good clustering performance, further proving the superiority of the scDRMAE method.

[0161] In addition, to further verify the ability of scDRMAE in learning dependencies between features and the impact of data interpolation mechanism on clustering, we performed cluster analysis on the data before and after interpolation. We conducted experiments on 6 multi-omics datasets. Figure 8 The clustering performance of Kmeans on imputed and unimputed data is shown. Figure 8 a and Figure 8 b. It can be seen that on different data sets, the clustering indicators after interpolation are better than those before interpolation, which confirms the influence of the interpolation mechanism on clustering performance. In order to more intuitively show the influence of the interpolation mechanism of scDRMAE on clustering, we draw the UMAP diagram of the original data and the reconstructed data, as shown in Figure 8 c and 8d. The left image shows the distribution of the original data, where many cells are mixed together and the boundaries of different cell types are unclear. The reconstructed data effectively separates different types of cells and clusters cells of the same type together. This phenomenon reflects the advantages of interpolation, making the distribution differences of different cell types more prominent, thereby improving the effect of cell clustering. In summary, the scDRMAE method shows good performance in data interpolation, and the use of interpolated omics data can effectively improve the clustering results.

[0162] Marker genes can be used to annotate specific cell types, and by studying their expression patterns, we can elucidate cellular heterogeneity and potential gene regulatory mechanisms. To further demonstrate the accuracy of the scDRMAE method for cell type division, we used a non-parametric Wilcoxon rank sum test to screen the top three differential features of each predicted cluster; 'gene, epiprotein, chromatin open region' and verified whether it could identify cell types through databases such as GeneCards and genebank. Fig. 9Figure a shows the top three differentially expressed genes in the cell clusters divided by scDRMAE on the InHouse dataset. We checked these differentially expressed genes in GeneCards and found that the differentially expressed genes screened from different clusters have biological significance and are strongly associated with a certain type of cell, further confirming the accuracy of scDRMAE cell type division.

[0163] For example, in cluster 0, CD8B was confirmed as a marker gene for CD8+ T cells, while IL7R was Marker genes for T cells and memory T cells. In cluster 1, the GNLY gene was confirmed to be a marker gene for natural killer cells (NK cells). The NKG7 gene is essential for the cytotoxic degranulation of NK cells and CD8 T cells, as well as the activation and proinflammatory response of CD4 T cells, and has also been proven to be a marker gene for NK cells. In addition, the granzyme protein encoded by the GZMB gene belongs to the member of the serine protease S1 family. Its precursor is secreted by NK cells and cytotoxic T lymphocytes (CTLs), and is processed by proteolysis to generate active proteases, inducing apoptosis of target cells. S100A8 and S100A9 found in cluster 2 belong to the calcium ion binding proteins of the S100 family. They are mainly expressed in neutrophils and monocytes, and play a key role in regulating various inflammatory responses and related diseases. Therefore, they can be used to identify monocytes. The marker genes CD79A and CD79B predicted in cluster 3 have been confirmed to be marker genes for B cells. The CD74 protein produced by the D74 gene is a type II transmembrane glycoprotein that is involved in regulating the survival signaling pathway of B cells. In addition, the B lymphocyte antigen CD20 is encoded by the MS4A1 gene. Since the CD20 protein is expressed at almost all stages of B cell development, it has become one of the most commonly used TIL-B biomarkers to date. The TRAC gene found in cluster 4 has been previously shown to be enriched in γδT cells. Finally, in cluster 5, CCL5 has been confirmed to play an important role in CD8+T cells, and NKG7 can also identify CD8+T cells to a certain extent.

[0164] We also performed a rank test analysis on the epiproteomics data of the InHouse dataset. Fig. 9b shows the top 3 marker proteins of different clusters screened in the Inhouse dataset. In cluster 0, CD8a was also confirmed as a marker for CD8+T cells. In cluster 1, CD56 was confirmed as a basic biomarker for natural killer cells 'NK cells'. In addition, NK cell subtypes can be further divided according to the relative expression levels of CD16 and CD56. In cluster 2, CD14 was identified as a marker for monocytes, which can send intracellular response signals when exposed to bacteria. CD11c is also highly expressed mainly in monocytes, further supporting the characteristics of this cell type. CD19 predicted in cluster 3 is a biomarker for normal and tumor B cells and follicular dendritic cells. CD19 plays a key role in forming the threshold of internal B cell signal transduction by regulating B cell receptor-dependent and independent signal transduction. In cluster 4, the discovery of CD4 protein highlights its importance as a marker for CD4+T cells. The CD4 protein is located on the surface of CD4+T cells, and molecules such as CD3, CD127 and CD154 are also highly expressed in T cells, further verifying the characteristics of T cells. In cluster 5, CD8a and CD154 have been proven to be biomarkers of T cells.

[0165] During the differential analysis, we found that although the two differentially expressed genes S100A8 and S100A9 in cluster 2 can not only identify monocytes but also neutrophils, if we use them to distinguish between the two cells, it is often difficult to separate them. However, we analyzed the protein data and screened the differentially expressed proteins CD14 and CD11c, which are marker proteins of monocytes. If we consider the two omics data comprehensively, we can easily distinguish them. This also explains why the use of multi-omics data can help us more accurately achieve cell type division. By performing differential analysis on the cell clusters divided by scDRMAE, we found that the top few differential features of each cluster have been confirmed to be biomarkers of a certain type of cell, which further confirms the accuracy of the cell type division of the scDRMAE model. In addition, although some of the biomarkers we screened have not been fully verified, through observation Fig. 9 a and 9b show that they are very different between different clusters. In order to show this difference more intuitively, we visualized these screened differential genes, which can be seen in 9c and 9d, indicating that these differential features may be potential biomarkers. Interestingly, we found that the differential protein CD127 screened from cluster 0 was transcribed and translated by the differential gene IL7R, indicating that the scDRMAE model has potential value and application prospects in exploring the association between different omics, and is expected to provide direction for biologists. The above examples confirm the excellent clustering performance of scDRMAE.

[0166] The above is only a preferred embodiment of the present invention and does not limit the present invention in any form. Although the present invention has been disclosed as a preferred embodiment as above, it is not used to limit the present invention. Any technician familiar with this profession can make some changes or modify the technical contents disclosed above into equivalent embodiments without departing from the scope of the technical solution of the present invention. However, any simple modification, equivalent replacement and improvement made to the above embodiments without departing from the content of the technical solution of the present invention, based on the technical essence of the present invention, within the spirit and principles of the present invention, still fall within the protection scope of the technical solution of the present invention.

[0167] It should be noted that in the above embodiments, as long as the technical solutions are not contradictory, they can be arranged and combined, and those skilled in the art can exhaust all possibilities based on the mathematical knowledge of arrangement and combination. Therefore, the present invention will no longer describe the technical solutions after arrangement and combination one by one, but it should be understood that the technical solutions after arrangement and combination have been disclosed by the present invention.

[0168] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present invention shall be included in the protection scope of the present invention.

Claims

1. A method for cell type classification based on single-cell multi-omics data, characterized in that: The steps include: Step 1: Build a deep neural network model; Step 2: Obtain a comprehensive single-cell multi-omics dataset; Step 3: Preprocessing the single-cell multi-omics dataset; Step 4: Train the deep neural network model based on the preprocessed single-cell multi-omics dataset; Step 5: Classify the cell types of the test data based on the trained deep neural network model.

2. A method for cell type classification based on single-cell multi-omics data according to claim 1, characterized in that: The deep neural network model consists of an encoder, a mask decoder, a decoder and an attention fusion module; The encoder, mask decoder, and decoder form a masked autoencoder; Different omics encoders are used to extract important features of different omics data and the dependencies between features; The mask decoder is used to decode the mask to predict the dropout event; The decoder is used to reconstruct and interpolate dropout events.

3. The method for cell type classification based on single-cell multi-omics data according to claim 1, characterized in that: The single-cell multi-omics comprehensive datasets include a multi-omics dataset consisting of scRNA-seq data and scATAC-seq data and a multi-omics dataset consisting of scRNA-seq data and scADT-seq data.

4. The method for cell type classification based on single-cell multi-omics data according to claim 1, characterized in that: The single-cell multi-omics dataset is preprocessed specifically including: Step 1: For scRNA-seq data, we filtered out genes expressed in less than 1% of cells, screened 3,000 highly variable genes, and performed log normalization and standardization; Step 2: For scATAC-seq data, remove the peaks present in less than 1% of cells, select 3000 highly variable peaks and perform binarization, and finally apply TF-IDF for feature transformation; Step 3: Repeat step 2 to perform the same log normalization and standardization on the scADT-seq data.

5. The method for cell type classification based on single-cell multi-omics data according to claim 1, characterized in that: The training steps for a deep neural network model include: Step 1: Generate mask data and randomly shuffle some features in the original data matrix; Step 2: Generate a mask matrix M whose items all obey the Bernoulli distribution RNA-seq , which is expressed as follows: M RNA-seq ~Bernoulli(P) (1) In formula (1), P represents the probability of dropout events occurring in different genes in different cells; Therefore, by analyzing the raw data X of scRNA-seq RNA-seq , disrupt data and the mask matrix M RNA-seq Combine to get the masked data matrix x RNA-seq , which is expressed as follows: Similarly, the mask data x of other omics others , which is expressed as follows: Step 3: Mask decoder and Low-dimensional representation Z for learning from different omics RNA-seq and Z others The mask matrix corresponding to different omics data is predicted in , which is expressed as follows: In formula (4), and Represents the encoders of different omics, M' RNA-seq and M' others is the mask matrix of different omics predicted by the mask decoder; By using binary cross entropy loss as the mask prediction loss L M , which enables the model to learn dropout events in omics data, as shown below: In formulas (6), (7), (8) and They represent the mask prediction losses of different omics, n represents the number of cells, g and p represent the feature dimensions of different omics data, and represents the real mask matrix, and represents the mask matrix predicted by the mask decoder; The predicted mask matrices and low-dimensional representations of different omics are combined and passed into the decoder for reconstruction of different omics data, as shown below: In formulas (8) and (9), X' RNA-seq and X' others They represent the reconstruction data of different omics data respectively; The reconstruction loss L is calculated by calculating the mean square error between the reconstructed data and the original data of different omics R , while assigning different weights to damaged and undamaged genes, as follows: and Represent the reconstruction loss of different omics data, and Represent the reconstruction weights of perturbed and unperturbed features of scRNA-seq data and other omics data, respectively; Step 4: Concatenate the low-dimensional representations of different omics, as shown below: From=[From RNA-seq ,WITH others ] (13) The feature transformation of Z is performed through the attention mechanism, so that Z can pay more attention to the features that are important to the clustering results; Using the same strategy as transformer, Z is mapped to three different feature subspaces: K=ZW1; Q=ZW2; V=ZW3 (14) In formula (14), W1, W2 and W3 represent three weight matrices used to map changes, and K, Q and V represent three feature codes after Z is linearly transformed; Use K and Q to calculate the global relationship matrix W: Where d represents the feature dimension of Z; With the help of W, V is enhanced to obtain the enhanced features. b is the deviation; By integrating the scRNA-seq data feature Z r Again and Fusion to obtain a more informative feature representation It is expressed as follows: The KL loss is introduced as a constraint, and the t distribution is used to describe the similarity between cells i and j: In formula (18) and represents the encoding representation of cells i and j, q ij represents the soft assignment, which is used to represent the pairwise similarity between cells i and j; p ij Based on q ij The target distribution is constructed by ij After normalization, the calculation process is as follows: By finding the KL divergence of the two distributions, the features are constrained to improve the clustering effect: Step 5: After obtaining the single-cell multi-omics fusion data, cluster analysis is performed on the single-cell multi-omics fusion data using the Kmeans algorithm.

Images