Single-cell multi-omics sequencing data integration method and system based on comparative learning
By using a contrastive learning-based approach, multiple sub-views are created and features are extracted using an autoencoder and multi-head attention mechanism. Combined with cross-modal attention and loss function optimization, the high sparsity and high noise problems in single-cell multi-omics sequencing data integration are solved, achieving efficient multi-omics feature fusion and robustness improvement.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-17
- Publication Date
- 2026-03-17
AI Technical Summary
Existing methods struggle to effectively integrate single-cell multi-omics sequencing data, especially in cases of high sparsity, high noise, and high dimensionality. They fail to fully explore local correlations and utilize potential correlations between unpaired multi-omics data, making it difficult for models to capture commonalities and dissimilarities among different omics.
We employ a contrastive learning-based approach, creating multiple subviews through a random masking mechanism. We use negative binomial zero-inflation loss to constrain the autoencoder to extract local and global features, combine a multi-head attention mechanism to merge local features, and design a mean squared error loss to constrain the consistency of global features. We utilize cross-modal attention to align cross-omics features and design multi-omics contrastive loss and matching loss to optimize the feature fusion process.
It improves the performance and robustness of single-cell multi-omics data fusion, reduces the impact of noise on the fusion results, and can effectively capture global-local feature dependencies within omics and correlations between omics data, generating high-quality cell characterization.
Smart Images
Figure CN121687191A_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the field of single-cell data analysis, in particular to a single-cell multi-omics sequencing data integration method and system based on contrast learning. BACKGROUND
[0002] In recent years, bioinformatics technology has developed rapidly under the driving of single-cell sequencing, artificial intelligence and big data analysis technology, and has promoted innovation breakthroughs in the fields of disease diagnosis and treatment and drug research and development. On the one hand, with the continuous progress of sequencing technology, scientists have gradually acquired the ability to quantitatively measure molecules from multiple omics such as genome, transcriptome, epigenetic group, etc. at the microscopic level. Specifically, single-cell nucleic acid ribose sequencing data (scRNA-seq) quantifies the gene expression level in the cell; single-cell chromatin transposase accessibility sequencing (scATAC-seq) provides the expression of the openness of the loss of regulatory elements in the adjacent gene. Joint analysis of these data from different omics can capture the mutual relationship between different omics and realize the complementary fusion of different omic information, so as to more specifically understand the cell function, cell state and gene regulation relationship, and provide valuable inspiration and support for further research on cell heterogeneity, biological development process and human diseases.
[0003] Although the progress of multi-omics sequencing technology provides more abundant data sources and more advanced data analysis methods, how to effectively integrate and analyze multi-source multi-omics biological information still faces a series of complex challenges. First, the single-cell sequencing data of different omics shows a cross relationship, and different omics data only describe part of the mechanism of the cell, and the omics information is related but not coincident; second, the single-cell multi-omics sequencing data has the characteristics of high sparsity, high noise and non-uniform dimension, which not only restricts the computer modeling analysis process of single-omics data, but also significantly increases the difficulty of fusion analysis of multi-omics data. Early multi-omics data fusion methods usually use non-negative matrix factorization or principal component analysis to map different omics data to the same space. However, the linear mapping method may cause the loss of omics-specific information and cannot effectively model the potential nonlinear relationship between multi-omics data. Therefore, manifold integration-based methods are proposed to realize the low-dimensional embedding alignment of different omics data. Although the manifold alignment method can solve the problem of nonlinear alignment, the computational overhead of multi-network construction and alignment increases exponentially when facing large amounts of data, which seriously limits its applicable scenarios. In recent years, deep learning technology has become the mainstream method for single-cell multi-omics data fusion due to its ability to capture feature relationships from complex high-dimensional data. Although a series of deep learning-based single-cell multi-omics fusion analysis methods have been proposed, they still have two core limitations: first, existing methods mainly focus on global macro correlations in single-omics sequencing data and fail to fully exploit and utilize the potential local correlation relationships in the data, making it difficult to effectively model high-dimensional and high-sparse single-omics data; second, existing methods fail to fully utilize the potential association between unpaired multi-omics data, making it difficult for the model to fully capture the common and unique information between different omics. Therefore, it is urgent to design a technical method that can overcome the above problems to better realize single-cell multi-omics sequencing data fusion. SUMMARY
[0004] The present application is to solve the above problems, and aims to provide a single-cell multi-omics sequencing data integration method and system based on contrast learning.
[0005] The present application provides a single-cell multi-omics sequencing data integration method based on contrast learning, which has the following characteristics: S1, creating multiple different sub-views for the single-omics data set through a random mask mechanism, obtaining transcriptome sub-view data and epigenetic group sub-view data ; S2, encoding the transcriptome sub-view data and epigenetic group sub-view data to obtain global features and local features; S3, merging the local features to obtain the merged single-omics multi-view features , ; S4, using mean square error loss to constrain consistency of global features and merged single-omic multi-view features , ; S5, aligning global features to obtain aligned global features; S6, optimizing alignment process in S5 based on multi-omic contrastive loss; S7, performing multi-omic feature fusion based on aligned global features to obtain multi-omic fusion features ; S8, optimizing fusion process of multi-omic fusion features based on multi-omic matching loss , wherein the global features include single-omic transcriptome global features and single-omic epigenetic group global features , the local features include single-omic transcriptome local features and single-omic epigenetic group local features , the aligned global features include aligned single-omic transcriptome global features and aligned single-omic epigenetic group global features .
[0006] In the single-cell multi-omic sequencing data integration method based on contrastive learning provided by the application, the following features can also be present: wherein S1 includes the following sub-steps: for single-cell transcriptome sub-view data and epigenetic group sub-view data in the single-cell multi-omic sequencing data set, randomly mask 40% of the column data, repeat 8 times to obtain transcriptome sub-view data and epigenetic group sub-view data , wherein represents the number of cells in the data set, represents the dimension of the transcriptome sequencing data, represents the dimension of the epigenetic group sequencing data.
[0007] In the single-cell multi-omic sequencing data integration method based on contrastive learning provided by the application, the following features can also be present: wherein S2 includes the following sub-steps: S2.1 based on single-omic data sets, transcriptome sub-view data and epigenetic group sub-view data , constructing single-omic auto-encoder and , wherein is a single-omic transcriptome auto-encoder, is a single-omic epigenetic group auto-encoder, and the encoding obtains: single-omic transcriptome global features , single-omic transcriptome local features , single-omic epigenetic group global features , and single-omic epigenetic group local features wherein is a learnable parameter, is a learnable parameter; S2.2 trains the omics autoencoder based on the single-omics transcriptome global features with a negative binomial zero-inflated ZINB reconstruction loss, the calculation formula of the reconstruction loss is:
[0008] (1),
[0009] wherein, is a negative binomial zero-inflated distribution, the calculation formula is as follows:
[0010] (2),
[0011] (3),
[0012] in the formula, represents the distribution mean, represents the distribution dispersion, represents the distribution zero inflation probability, , respectively represent the matrix form of in the transcriptome sub-view data and the epigenetic group sub-view data, , respectively represent the matrix form of in the transcriptome sub-view data and the epigenetic group sub-view data, , respectively represent the matrix form of in the transcriptome sub-view data and the epigenetic group sub-view data, the calculation formulas are respectively:
[0013] (4),
[0014] (5),
[0015] (6),
[0016] in the formula, are independent decoders corresponding to the transcriptome and the epigenetic group respectively, , and are learnable parameters belonging to three different fully connected layers in , represents a sigmoid activation function, represents an exponential function.
[0017] In the single-cell multi-omics sequencing data integration method based on contrast learning provided by the application, the S3 can further include the following sub-steps: based on the single-omics transcriptome global feature , the single-omics transcriptome local feature , the single-omics epigenetic group global feature and the single-omics epigenetic group local feature , the single-omics transcriptome local feature and the single-omics epigenetic group local feature are combined using a multi-head attention mechanism, the local feature representation in each view is learned uniformly, and the calculation formula is:
[0018] (7)
[0019] (8)
[0020] (9)
[0021] (10)
[0022] In the formula, all of , , and are learnable parameters, represents a vector splicing operation, represents the number of attention heads, is the feature dimension of each attention head, represents the dimension of the local feature , and is the combined single-omics multi-view feature.
[0023] In the single-cell multi-omics sequencing data integration method based on contrast learning provided by the application, the S4 can further include the following features: in the S4, a soft fusion strategy is designed, and a mean square error loss is used to constrain the single-omics transcriptome global feature , the single-omics epigenetic group global feature and the combined single-omics multi-view feature , , so as to unify the global-local feature representation, and the calculation formula is:
[0024] (11).
[0025] In the single-cell multi-omics sequencing data integration method based on contrast learning provided by the application, the S5 can further include the following features: in the S5, the single-omics transcriptome global feature and the single-omics epigenetic group global feature The aligned global features of the single-omics transcriptome were obtained. Aligned single-omics epigenome global features The calculation formula is:
[0026] (12)
[0027] (13)
[0028] In the formula , and All of these are learnable parameters.
[0029] The single-cell multi-omics sequencing data integration method based on contrastive learning provided in this invention may also have the following feature: In S6, multi-omics contrastive loss is used to supervise the alignment process in S5, and the multi-omics contrastive loss is applied to each input... Each cell is considered an independent batch, and the calculation formula is as follows:
[0030] (14)
[0031] (15)
[0032] In the formula For true unique heat encoding tags, their shape is as follows: The identity matrix, This is a similarity matrix from transcriptome to epigenome. This is a similarity matrix from the epigenome to the transcriptome. For the first Transcriptome features and the first Cosine similarity of epigenetic features.
[0033] The single-cell multi-omics sequencing data integration method based on contrastive learning provided in this invention may also have the following feature: wherein, in S7, based on the aligned global features of the single-omics transcriptome... Aligned single-omics epigenome global features Using multi-head sensor Achieving multi-omics feature fusion to obtain multi-omics fused features The calculation formula is:
[0034] (16)
[0035] In the formula: For feature splicing operations, These are learnable parameters.
[0036] The single-cell multi-omics sequencing data integration method based on contrastive learning provided in this invention may also have the following feature: wherein, in S8, based on the multi-omics data matching relationship, negative samples are generated by splicing and fusing unpaired data. At the same time, it integrates the characteristics of multiple omics. As positive samples, a multi-omics matching loss is introduced to supervise the multi-omics feature fusion process. The calculation formula is as follows:
[0037] (17)
[0038] (18)
[0039] In the formula: It is a fully connected layer. Its learnable parameters, These are one-hot encoded labels representing positive and negative samples.
[0040] This invention also provides a single-cell multi-omics sequencing data integration system based on contrastive learning, characterized by including: a subview creation module that creates multiple different subviews for a single-omics dataset through a random masking mechanism to obtain transcriptome subview data. Epigenome subview data Feature encoding module, encoding transcriptome subview data. Epigenome subview data The module obtains global and local features; the local feature merging module merges the local features to obtain the merged single-omics multi-view features. , The consistency constraint module uses mean squared error loss to constrain global features and merged single-omics multi-view features. , The system includes: a consistency module; a global feature alignment module to align global features and obtain aligned global features; an alignment optimization module to optimize the alignment process in S5 based on multi-omics contrastive loss; and a multi-omics feature fusion module to perform multi-omics feature fusion based on the aligned global features and obtain multi-omics fused features. The fusion optimization module optimizes multi-omics fusion features based on multi-omics matching loss. The fusion process, in which global features include single-omics transcriptome global features. and global epigenetic features of single omics Local features include single-omics transcriptome local features. Local characteristics of the epigenome in single omics Aligned global features include aligned single-omics transcriptome global features. Aligned single-omics epigenome global features .
[0041] The role and effect of invention
[0042] This invention relates to a method and system for integrating single-cell multi-omics sequencing data based on contrastive learning. Addressing the challenge of multi-omics data fusion caused by high noise, high sparsity, and high dimensionality differences in single-cell multi-omics sequencing data, this invention proposes a method that effectively captures global-local feature dependencies within an omics dataset while simultaneously mining correlations between different omics datasets. The invention proposes an intra-omics contrastive learning method, creating multiple different sub-views for a single omics sequencing data dataset using a random masking mechanism. Local and global features are encoded using an autoencoder based on a negative binomial zero-inflation loss constraint. Simultaneously, a multi-head attention mechanism is used to merge local features. Finally, a soft fusion strategy based on a mean squared error loss constraint is designed to capture the potential local and global interactions within a single omics dataset. In the multi-omics feature fusion stage, this invention proposes an inter-omics contrastive learning method, achieving cross-omics feature interaction alignment based on cross-modal attention. Multi-omics contrastive loss and multi-omics matching loss are designed to optimize the feature alignment and feature fusion processes, respectively. These methods improve the performance and robustness of this invention in the single-cell multi-omics data fusion process and reduce the impact of noise in the data on the fusion results. Attached Figure Description
[0043] Figure 1 This is an algorithm framework diagram of the single-cell multi-omics sequencing data integration method based on contrastive learning in an embodiment of the present invention.
[0044] Figure 2 This is a comparison of cell type clustering performance of single-cell multi-omics data fusion methods on four datasets in the embodiments of the present invention.
[0045] Figure 3 These are the cell subtype analysis results on the PBMC-10K dataset in the embodiments of the present invention.
[0046] Figure 4 These are the results of cell development trajectory inference experiments on the PBMC-10K dataset in the embodiments of the present invention.
[0047] Figure 5 This is a schematic diagram of a single-cell multi-omics sequencing data integration system based on contrastive learning in an embodiment of the present invention. Detailed Implementation
[0049] To make the technical means, creative features, objectives and effects of this invention easier to understand, the following embodiments, in conjunction with the accompanying drawings, specifically illustrate the method and system for integrating single-cell multi-omics sequencing data based on contrastive learning.
[0050] Figure 1 This is an algorithm framework diagram of the single-cell multi-omics sequencing data integration method based on contrastive learning in an embodiment of the present invention.
[0051] like Figure 1 As shown, this invention discloses a method for integrating single-cell multi-omics sequencing data based on contrastive learning, including single-omics feature extraction, multi-omics feature fusion, and downstream task implementation, specifically including the following steps:
[0052] S1 uses a random masking mechanism to create multiple different subviews for a single-omics dataset, resulting in transcriptome subview data. Epigenome subview data .
[0053] S1 includes the following sub-steps:
[0054] Single-cell transcriptome subview data in single-cell multi-omics sequencing datasets and epigenetic subview data Randomly mask 40% of the data, repeat 8 times, to obtain transcriptome subview data. Epigenome subview data .
[0055] in Represents the number of cells in the dataset. Represents the dimensions of transcriptome sequencing data, This represents a dimension of epigenetic sequencing data.
[0056] S2, encoding transcriptome subview data Epigenome subview data This yields global and local features.
[0057] S2 includes the following sub-steps:
[0058] S2.1 is based on single-omics datasets and transcriptome subview data. and epigenome subview data Constructing a single-omics autoencoder and ,in It is a single-omics transcriptome autoencoder. The single-omics epigenome autoencoder encoded the following:
[0059] Global features of single-omics transcriptome .
[0060] Local features of single-omics transcriptome .
[0061] Global epigenomic characteristics of single omics .
[0062] Local characteristics of a single-omics epigenetic genome .
[0063] in for Learnable parameters for Learnable parameters.
[0064] S2.2 Based on global features of single-omics transcriptome A single-omics autoencoder is trained using the negative binomial zero-inflated ZINB reconstruction loss. The formula for calculating the reconstruction loss is as follows:
[0065] (1).
[0066] in, For a negative binomial zero-inflated distribution, its calculation formula is as follows:
[0067] (2).
[0068] (3).
[0069] In the formula Represents the mean of the distribution. Represents the dispersion of the distribution. Represents the probability of zero inflation in the distribution. , These represent the transcriptome subview data and the epigenome subview data, respectively. In matrix form, , These represent the transcriptome subview data and the epigenome subview data, respectively. In matrix form, , These represent the transcriptome subview data and the epigenome subview data, respectively. The matrix form and the calculation formulas are as follows:
[0070] (4).
[0071] (5).
[0072] (6).
[0073] In the formula These are independent decoders corresponding to the transcriptome and epigenome, respectively. , and They are respectively The learnable parameters of the three different fully connected layers in the middle. This represents the sigmoid activation function. This represents an exponential function.
[0074] S3, merge local features to obtain merged single-omics multi-view features. , .
[0075] S3 includes the following sub-steps:
[0076] Based on global features of single-omics transcriptome Local features of single-omics transcriptome Global characteristics of single-omics epigenome and local features of single-omics epigenetics Using multi-head attention mechanisms to merge local features of single-omics transcriptomes Local characteristics of the epigenome in single omics The local feature representations in each view are uniformly learned, and the calculation formula is as follows:
[0077] (7).
[0078] (8).
[0079] (9).
[0080] (10).
[0081] In the formula , , and All are learnable parameters. This represents a vector concatenation operation. Represents the number of attention heads, For each attention head, the feature dimensions, Representing local features Dimensions This represents the merged single-omics multi-view features.
[0082] S4, using mean squared error loss to constrain global features and merged single-omics multi-view features. , Consistency.
[0083] In S4: A soft fusion strategy was designed, using mean squared error loss to constrain global features of single-omics transcriptomes. Global epigenetic characteristics of single omics and the merged single-omics multi-view features , This unifies the global-local feature representation, and the calculation formula is:
[0084] (11)
[0085] S5, align global features to obtain aligned global features.
[0086] In S5, cross-modal attention is used to align global features of a single-omics transcriptome. and global epigenetic features of single omics The aligned global features of the single-omics transcriptome were obtained. Aligned single-omics epigenome global features The calculation formula is:
[0087] (12).
[0088] (13).
[0089] In the formula , and All of these are learnable parameters.
[0090] S6 optimizes the alignment process in S5 based on multi-omics contrastive loss.
[0091] In S6, a multi-omics contrastive loss is used to supervise the alignment process in S5. The multi-omics contrastive loss is applied to each input. Each cell is considered an independent batch, and the calculation formula is as follows:
[0092] (14).
[0093] (15).
[0094] In the formula For true unique heat encoding tags, their shape is as follows: The identity matrix, This is a similarity matrix from transcriptome to epigenome. This is a similarity matrix from the epigenome to the transcriptome. For the first Transcriptome features and the first Cosine similarity of epigenetic features.
[0095] S7, based on the aligned global features, performs multi-omics feature fusion to obtain multi-omics fused features. .
[0096] In S7, based on aligned single-omics transcriptome global features Aligned single-omics epigenome global features Using multi-head sensor Achieving multi-omics feature fusion to obtain multi-omics fused features The calculation formula is:
[0097] (16).
[0098] In the formula: For feature splicing operations, These are learnable parameters.
[0099] S8, Optimizing Multi-Omics Fusion Features Based on Multi-Omics Matching Loss The fusion process.
[0100] In S8, negative samples are generated by splicing and fusing unpaired data based on multi-omics data matching relationships. At the same time, it integrates the characteristics of multiple omics. As positive samples, a multi-omics matching loss is introduced to supervise the multi-omics feature fusion process. The calculation formula is as follows:
[0101] (17).
[0102] (18).
[0103] In the formula: It is a fully connected layer. Its learnable parameters, These are one-hot encoded labels representing positive and negative samples.
[0104] Verification experiment:
[0105] Four pairs of real-world single-cell multi-omics sequencing datasets were selected, including the 10x-Multiome-PBMC 10k (PBMC-10K) dataset, the 10x-Multiome-PBMC 3k (PBMC-3K) dataset, the mouse epidermal cell (Ma-2020) dataset, and the human mixed cell (CellMix) dataset. The transcriptome sequencing data matrix and the epigeneome sequencing data matrix from these datasets were used as inputs to the model of this invention to obtain the fused cell representation.
[0106] The four selected real-world single-cell multi-omics sequencing datasets are:
[0107] The PBMC-10K dataset (from https: / / support.10xgenomics.com / single-cell-multiome-atac-gex / datasets / 1.0.0 / pbmc_granulocyte_sorted_10k) contains 9631 cells, divided into 19 classes. The transcriptome sequencing data has a dimension of 29095, and the epigenetics sequencing data has a dimension of 107194.
[0108] The PBMC-3K dataset (from https: / / www.10xgenomics.com / datasets) contains 2585 cells divided into 14 classes, with the transcriptome sequencing data having a dimension of 36601 and the epigenetics sequencing data having a dimension of 20010.
[0109] The Ma-2020 dataset (excerpted from Ma S, Zhang B, LaFave LM, et al. Chromatinpotential identified by shared single-cell profiling of RNA and chromatin[J].Cell, 2020, 183(4): 1103-1116. e20.) contains 32,231 cells, divided into 22 categories. The transcriptome sequencing data has a dimension of 21,478, and the epigenetics sequencing data has a dimension of 340,341.
[0110] The CellMix dataset (excerpted from Chen S, Lake BB, Zhang K. High-throughput sequencing of the transcriptome and chromatin accessibility in the same cell[J]. Nature biotechnology, 2019, 37(12): 1452-1457.) contains 1047 cells, divided into 4 categories, with the transcriptome sequencing data having a dimension of 18666 and the epigenetics sequencing data having a dimension of 136771.
[0111] To verify the performance of the method of the present invention, the verification experiments of the present invention include: (1) performance comparison of cell type clustering; (2) the impact of the strategy proposed in the present invention on the clustering results; and (3) cell development trajectory reasoning analysis.
[0112] The specific experimental setup is as follows:
[0113] Cell type clustering experiments used ARI and NMI as evaluation metrics. The mini-batch size was set to 128, the learning rate to 0.001, and the Adam optimizer was used as the optimizer for the entire experiment.
[0114] (1) Performance comparison of cell type clustering
[0115] Eight different single-cell data analysis methods were used, including five single-cell multi-omics data integration methods (SCMLC, DCCA, scMCs, VIMCCA, and scMVAE), two single-cell transcriptome data analysis methods (scGPT and scVI), and one single-cell epigenetics data analysis method (PeakVI) as comparison methods. The models were trained on four datasets respectively.
[0116] Figure 2 This is a comparison of cell type clustering performance of single-cell multi-omics data fusion methods on four datasets in the embodiments of the present invention.
[0117] Experimental results are as follows Figure 2 As shown in the experimental results, the clustering results on four real-world datasets outperformed other comparative methods in both evaluation metrics, demonstrating that this method can integrate and generate high-quality cell representations. This is further demonstrated by introducing the rank score metric. ,in To compare the number of methods, The average ranking of the method on four different datasets shows that the present invention has achieved a significant lead in global experiments, indicating that the method can achieve excellent performance on different species and data of different sizes, further demonstrating the robustness of the method.
[0118] (2) The impact of the strategy proposed in this invention on the clustering results:
[0119] w / o CM: Remove the multi-omics alignment module and directly use single-omics features for multi-omics data fusion.
[0120] w / o Rec: Removes the negative binary zero-inflation reconstruction loss term.
[0121] w / o MOC: Remove multi-omics contrastive loss term.
[0122] w / o MOM: Removes the multi-omics matching loss term.
[0123] w / o Intra-CL: Remove the intra-group contrastive learning module and directly use the autoencoder to extract single-omics features from the sequencing data;
[0124] w / o Inter-CL: Removes the inter-group contrastive learning module and uses only a single multilayer perceptron to achieve multi-omics feature fusion.
[0125] Figure 3 These are the cell subtype analysis results on the PBMC-10K dataset in the embodiments of the present invention.
[0126] Experimental results are as follows Figure 3 As shown in the experimental results, the present invention outperforms the aforementioned variant methods, verifying the effectiveness of each component of the system. In the method of the present invention, removing Intra-CL (intra-group contrastive learning) and Inter-CL (inter-group contrastive learning module) both lead to a significant performance decrease, indicating that both play crucial roles in multi-omics data fusion.
[0127] Specifically, removing the inter-group contrastive learning module significantly reduced the performance of the variant method, causing its model performance to fall below that of most contrastive methods. This demonstrates that the contrastive learning strategy plays a crucial role in learning high-quality cell representations. Furthermore, removing any single module or loss term from the inter-group contrastive learning process alone did not result in a similar performance loss, indicating that the effectiveness of the inter-group contrastive learning strategy stems from the combined contributions of all modules.
[0128] (3) Cell development trajectory reasoning analysis
[0129] Three different single-cell data analysis methods were used as comparison methods: VIMCCA, a single-cell multi-omics data integration method; a single-cell transcriptome data analysis method based on principal component analysis (PCA) dimensionality reduction; and a single-cell epigenetics data analysis method based on latent semantic search (LSI). Cell development trajectory inference analysis experiments were conducted on the PBMC-10K dataset.
[0130] Figure 4 These are the results of cell development trajectory inference experiments on the PBMC-10K dataset in the embodiments of the present invention.
[0131] Experimental results are as follows Figure 4 As shown, from the experimental results, only this invention can be used to deduce HSPC. Naive Bcell Intermediate B cells The correct cell development path of Memory B cells further demonstrates that the method of the present invention has the ability to generate high-quality fusion cell characterization.
[0132] Based on the steps of the above method, this invention includes single-omics feature extraction, multi-omics feature fusion, and downstream task implementation:
[0133] I. Single-atom feature extraction, including:
[0134] Multiple distinct sub-views of single-cell transcriptome sequencing data and single-cell epigenetics sequencing data are generated using a random masking mechanism.
[0135] An autoencoder is constructed to extract single-omics global and local features from single-cell sequencing data and its subviews, respectively. Based on the single-omics global features, the single-omics autoencoder is trained using negative binomial zero-inflation reconstruction loss. Multi-view feature spaces are merged based on multi-head attention to obtain merged local features. A soft fusion strategy is designed to unify the global-local feature representations through mean squared error loss constraints.
[0136] II. Fusion of multi-omics features, including:
[0137] Based on cross-modal attention, the heterogeneous information between omics is preserved while achieving interactive alignment of multi-omics features; positive and negative sample pairs are created based on multi-omics pairing relationships, and the feature alignment process is optimized using inter-omics contrast loss; paired multi-omics features are fused to obtain cell-level representations; positive and negative sample pairs are created based on multi-omics feature matching relationships, and the feature fusion process is optimized using inter-omics matching loss.
[0138] III. In the downstream task implementation stage:
[0139] Cell type clustering is achieved using the K-means algorithm, and feature dimensionality reduction and visualization are implemented using the UMAP algorithm. Cell subtype analysis and cell type relabeling are achieved using hypervariable gene analysis and gene set enrichment analysis algorithms. Cell development trajectory inference is achieved using the Slingshot algorithm. This invention implements the method on a real-world single-cell multi-omics sequencing dataset, demonstrating that the method can effectively learn high-quality cell characterizations and provide effective support for downstream cell analysis.
[0140] Figure 5 This is a schematic diagram of a single-cell multi-omics sequencing data integration system based on contrastive learning in an embodiment of the present invention.
[0141] like Figure 5 As shown, this invention also discloses a single-cell multi-omics sequencing data integration system based on contrastive learning, obtained according to the above-described single-cell multi-omics sequencing data integration method based on contrastive learning, comprising:
[0142] The subview creation module, using the method described in S1 above, creates multiple different subviews for a single-omics dataset through a random masking mechanism, resulting in transcriptome subview data. Epigenome subview data .
[0143] The feature encoding module, obtained using the method described in S2 above, encodes transcriptome subview data. Epigenome subview data This yields global and local features.
[0144] The local feature merging module, using the method described in S3 above, merges local features to obtain the merged single-omics multi-view features. , .
[0145] The consistency constraint module is obtained using the method described in S4 above, employing mean squared error loss to constrain global features and merged single-omics multi-view features. , Consistency.
[0146] The global feature alignment module uses the method described in S5 above to align global features, resulting in aligned global features.
[0147] The alignment optimization module is obtained by using the method in S6 above, and the alignment process in S5 is optimized based on multi-omics contrastive loss.
[0148] The multi-omics feature fusion module, obtained using the method described in S7 above, performs multi-omics feature fusion based on the aligned global features to obtain multi-omics fused features. .
[0149] The fusion optimization module is obtained using the method described in S8 above, which optimizes multi-omics fusion features based on multi-omics matching loss. The fusion process.
[0150] Among them, global features include single-omics transcriptome global features. and global epigenetic features of single omics Local features include single-omics transcriptome local features. Local characteristics of the epigenome in single omics .
[0151] Aligned global features include aligned single-omics transcriptome global features. Aligned single-omics epigenome global features .
[0152] Global features include single-omics transcriptome global features. and global epigenetic features of single omics Local features include single-omics transcriptome local features. Local characteristics of the epigenome in single omics .
[0153] Aligned global features include aligned single-omics transcriptome global features. Aligned single-omics epigenome global features .
[0154] The role and effect of the embodiments
[0155] Those skilled in the art should understand that this invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Various changes and modifications can be made to this invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the invention as claimed. The scope of protection of this invention is defined by the appended claims and their equivalents.
Claims
1. A method for integrating single-cell multi-omics sequencing data based on contrastive learning, characterized in that, comprising the following steps: S1, creating multiple different sub-views for a single omics dataset by a random masking mechanism, obtaining transcriptomic sub-view data with epigenetic group sub-view data ; S2, encoding the transcriptomic subview data with the epigenetic subview data , obtaining global features and local features; S3, merging the local features to obtain a merged single-omic multi-view feature , ; S4, using a mean squared error loss to constrain the global feature and the merged single-omic multi-view feature , consistency. S5, aligning the global features to obtain aligned global features; S6, optimizing the alignment process in S5 based on multi-omics contrast loss; S7, performing multi-omics feature fusion based on the aligned global features to obtain multi-omics fusion features ; S8, optimizing the multi-omics fusion features based on multi-omics matching loss of the fusion process, wherein the global features comprise single-omic transcriptome global features and single-omic epigenetic panel global features , the local features comprise single-omic transcriptome local features and the single-omic epigenetic panel local features , the aligned global features include aligned single-omic transcriptome global features and aligned single-omic epigenetic panel global features .
2. The single-cell multi-omics sequencing data integration method based on contrast learning according to claim 1, characterized in that: wherein S1 comprises the following sub-steps: For the single cell transcriptomic sub-views data in the single cell multi-omics sequencing dataset and epigenetic sub-views data , randomly mask 40% of the column data, repeat 8 times, get transcriptomic sub-views data and epigenetic sub-views data , wherein representing the number of cells in the dataset, representing the dimension of the transcriptome sequencing data, representing the dimension of the epigenome sequencing data.
3. The single-cell multi-omics sequencing data integration method based on contrast learning according to claim 1, characterized in that: wherein S2 comprises the following sub-steps: S2.1 constructing a single-omic autoencoder based on the single-omic dataset, the transcriptomic subview data and the epigenomic subview data wherein is a single-omic transcriptomic autoencoder, is a single-omic epigenomic autoencoder, encoding yields: The single-omic transcriptomic global features , The single-omic transcriptome local feature , The single-omic epigenetic panel global features , The single-omic epigenetic panel local features , wherein is a learnable parameter, is a learnable parameter; S2.2 Based on the global features of the single-omics transcriptome The single-omics autoencoder is trained using the negative binomial zero-inflated ZINB reconstruction loss. The formula for calculating the reconstruction loss is as follows: （1）， wherein, is a negative binomial zero-inflated distribution, which is calculated as follows: （2）， （3）， wherein represents the distribution mean, represents the distribution dispersion, represents the distribution zero inflation probability, , respectively represent the matrix form of the transcriptional subview data and the epigenetic subview data in , , respectively represent the matrix form of the transcriptional subview data and the epigenetic subview data in , , respectively represent the matrix form of the transcriptional subview data and the epigenetic subview data in , and the calculation formulas are respectively （4）， （5）， （6）， wherein are independent decoders for the transcriptomic and epigenetic groups, respectively, , and are are the learnable parameters of the three different fully connected layers in denotes the sigmoid activation function, denotes the exponential function.
4. The single-cell multi-omics sequencing data integration method based on contrast learning according to claim 3, characterized in that: wherein, S3 comprises the following sub-steps: based on the single-omic transcriptome global feature the single-omic transcriptome local feature based on the single-omic epigenetic panel global feature and the single-omic epigenetic panel local feature combining the single-omic transcriptome local feature and the single-omic epigenetic panel local feature using a multi-head attention mechanism, and learning local feature representations in each view uniformly, with the formula: （7）， （8）， （9）， （10）， wherein , , and are learnable parameters, denotes a vector concatenation operation, denotes the number of attention heads, is the feature dimension for each attention head, denotes the dimension of local features , is the merged omics multi-view feature.
5. The single-cell multi-omics sequencing data integration method based on contrast learning according to claim 1, characterized in that: wherein In S4: designing a soft fusion strategy using a mean squared error loss to constrain the global features of the single-omic transcriptome the global features of the single-omic epigenetic panel and the merged single-omic multi-view features , to unify the global-local feature representation, with the formula being: （11）。 6. The single-cell multi-omics sequencing data integration method based on contrast learning according to claim 1, characterized in that: wherein, In S5, the single-omic transcriptome global feature is aligned using cross-modal attention and the single-omic epigenetic panel global feature to obtain the aligned single-omic transcriptome global feature and the aligned single-omic epigenetic panel global feature The calculation formula is: （12）， （13）， wherein , and are learnable parameters.
7. The single-cell multi-omics sequencing data integration method based on contrast learning according to claim 1, characterized in that: wherein In the S6, a multi-omics contrast loss is used to supervise the alignment process of the S5, the multi-omics contrast loss inputs one cell as an independent batch each time, and the calculation formula is as follows: The formula is as follows: （14）， （15）， wherein is a unit matrix of the form is a unit matrix of the form is a similarity matrix of the transcriptome to the epigenome, is a similarity matrix of the epigenome to the transcriptome, is the cosine similarity between the th transcriptomic feature and the th epigenomic feature.
8. The single-cell multi-omics sequencing data integration method based on contrast learning according to claim 1, characterized in that: wherein, In the S7, based on the aligned single-omic transcriptome global feature and the aligned single-omic epigenetic group global feature , using a multi-head perception machine Implement multi-omic feature fusion to obtain the multi-omic fusion feature The calculation formula is: （16）， In the formula: is a feature stitching operation, are learnable parameters.
9. The single-cell multi-omics sequencing data integration method based on contrast learning according to claim 1, characterized in that: wherein, In S8, based on the multi-omics data matching relationship, the non-paired data is spliced and fused to generate negative samples At the same time, the multi-omics fusion features As a positive sample, a multi-omics matching loss is introduced to supervise the process of multi-omics feature fusion, and the calculation formula is: （17）， （18）， In the formula: is a fully connected layer, is a learnable parameter thereof, is a one-hot encoded label representing positive and negative samples.
10. A contrastive learning based single-cell multi-omics sequencing data integration system, characterized in that, comprising: A sub-view creation module creates multiple different sub-views for a single omics dataset through a random masking mechanism to obtain transcriptome sub-view data and epigenetic group sub-view data ; a feature encoding module to encode the transcriptomic subview data with the epigenetic group subview data to obtain global features and local features; a local feature merging module configured to merge the local features to obtain merged single-omic multi-view features , ; a consistency constraint module that constrains the global feature and the merged omics multi-view feature using a mean squared error loss , consistency a global feature alignment module, configured to align the global features to obtain aligned global features; an alignment optimization module, configured to optimize the alignment process in S5 based on multi-omics contrast loss; The multi-omics feature fusion module fuses the multi-omics features based on the aligned global features to obtain multi-omics fusion features ; a fusion optimization module that optimizes the multi-omics fused features based on a multi-omics matching loss of the fusion process, wherein the global features comprise single-omic transcriptome global features and single-omic epigenetic panel global features , the local features comprise single-omic transcriptome local features and the single-omic epigenetic panel local features , the aligned global features include aligned single-omic transcriptome global features and aligned single-omic epigenetic panel global features .