Single cell space transcriptome cell localization and type deconvolution algorithm based on feature decoupling
By combining single-cell RNA sequencing and spatial transcriptomics data with a cross-domain feature decoupling framework, the problem of not being able to fully utilize the complementarity of the two in existing technologies is solved, achieving high-resolution cell spatial mapping and type deconvolution, thereby improving the model's prediction accuracy and adaptability.
Patent Information
- Application Number
- CN202510764218.2
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-09
- Publication Date
- 2025-10-31
AI Technical Summary
Existing single-cell RNA sequencing and spatial transcriptomics technologies cannot fully utilize their complementarity when revealing cellular heterogeneity, resulting in loss of cellular spatial information or low resolution. Current methods have failed to effectively combine single-cell RNA sequencing and spatial transcriptomics data to improve the performance of both tasks.
A cross-domain feature decoupling framework is adopted, which extracts cross-modal shared features through a domain separation network and ensures feature decoupling through orthogonal constraints. The framework integrates single-cell pseudo-space reconstruction and cell type deconvolution of spatial transcriptomics probe points to construct a multi-task learning framework to achieve mutual supervision between the two modes.
It significantly improves the interpretability and application potential of single-cell RNA sequencing and spatial transcriptomics data, enhances the model's generalization ability to unknown cell spaces and cell type compositions, and improves prediction accuracy and robustness, especially showing excellent performance in complex cancer datasets.
Smart Images
Figure CN120877871A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of deep learning technology, specifically to an application of deep learning technology in revealing cellular heterogeneity, and more particularly to a single-cell spatial transcriptome cell localization and type deconvolution algorithm based on feature decoupling. Background Technology
[0002] Single-cell RNA sequencing, capable of unbiased analysis of transcriptional levels in individual cells, has become the standard method for assessing cellular composition in complex tissues. While single-cell RNA sequencing offers opportunities to reveal the heterogeneity of complex tissues, it requires dissociating the tissue into single-cell suspensions, leading to a loss of cellular spatial information. The spatial localization of cells plays a crucial role in identifying cellular state and function within complex tissues. Spatial transcriptomics has become a powerful tool for capturing spatial patterns of gene expression in diseased tissues.
[0003] Current spatial transcriptomics technologies can be broadly categorized into two types: (i) imaging-based protocols, such as MERFISH and seqFISH+, and (ii) sequencing-based protocols, such as STARmap and 10x Visium. However, imaging-based protocols often have a limited number of genes that can be analyzed in a single assay, while sequencing-based protocols like 10x Visium aggregate multiple cells into a single probe site, resulting in lower gene detection sensitivity.
[0004] In recent years, the complementary advantages of single-cell RNA sequencing and spatial transcriptomics have attracted significant attention in the research community. Single-cell RNA sequencing provides high-resolution, genome-wide expression profiles at the single-cell level, but spatial information is lost during tissue dissection. Conversely, spatial transcriptomics preserves the spatial arrangement of cells in tissue sections, but its resolution is lower because each "probe point" typically aggregates transcripts from multiple neighboring cells. To leverage the complementarity of these two approaches, several methods have been developed to use precise cell type markers obtained from single-cell RNA sequencing to guide cell type unconvolution in spatial transcriptomics data.
[0005] CARD is a spatially informed deconvolution method that combines single-cell RNA sequencing data with spatial transcriptomics. It improves deconvolution accuracy and achieves high-resolution spatial mapping by modeling the spatial correlations of cell type composition across different tissue locations. On the other hand, only a few methods have been proposed for inferring the spatial location of single cells, known as pseudo-spatial reconstruction. For example, scSpace uses transfer component analysis to reconstruct spatial associations of cells and leverages spatial transcriptomics data as a reference to depict spatially heterogeneous subpopulations from single-cell data. Similarly, Tangram aims to learn spatially resolved single-cell transcriptomes and address the inherent deconvolution challenges of spatial transcriptomics data. However, currently, no method fully leverages the complementarity of these two models to simultaneously improve performance on both tasks.
[0006] In this invention, we propose a cross-domain feature untangling framework that surpasses existing state-of-the-art benchmarks on two tasks, enhancing the interpretability and application potential of single-cell RNA sequencing and spatial transcriptomics data. Summary of the Invention
[0007] This invention aims to address the shortcomings of existing technologies by employing a domain separation network to extract cross-modal shared features. The framework can simultaneously incorporate cell type features and spatial location features, and by enforcing orthogonal constraints, it ensures that shared features are decoupled from interfering factors specific to each modality. This method integrates pseudo-spatial reconstruction of single cells with cell type deconvolution of spatial transcriptomics probe points into a multi-task learning framework, thereby achieving mutual supervision between the two modalities and fully utilizing their complementary information.
[0008] The present invention utilizes cross-domain transfer learning technology to construct a deep learning model to predict cell type composition in single-cell pseudo-space and spatial transcriptome. 1) Single-cell transcriptome data and spatial transcriptome data with all genes expressing zeros, and all genes expressing zeros are screened. Highly variable genes are then selected from paired single-cell transcriptome and spatial transcriptome data. 2) The data is standardized, and the union of the top 2000 highly variable genes from each paired single-cell transcriptome and spatial transcriptome data is retained. 3) Paired single-cell transcriptome and spatial transcriptome data are input into the model. A common feature encoder decouples the corresponding shared features, and a domain-specific feature encoder decouples the corresponding domain-specific features. To achieve complete feature decoupling, orthogonal constraints are applied to the shared features and domain-specific features to ensure their independence. 4) Shared features from the single-cell domain are input into the cell type predictor. Cross-entropy loss constraints are applied to the predicted cell types and the actual cell types to ensure that the shared features contain cell type information. 5) The shared features of the spatial transcriptome domain are input into the spatial location predictor. Mean squared error loss constraints are applied to both the predicted and actual spatial coordinates to ensure that the shared features contain spatial location information. 6) The shared features and domain-specific features are concatenated, and the domain input features are reconstructed using a shared decoder. Reconstruction loss is applied as a constraint, and the entire system is trained to ensure that the domain-specific features retain the decoupled information. 7) The trained common encoder is used to extract the shared features of the two domains. The shared features of the single-cell domain are input into the spatial location predictor to obtain the pseudo-spatial coordinates of a single cell. The shared features of the spatial transcriptome domain are input into the cell type predictor to obtain the cell type prediction score of the spatial transcriptome. This prediction score is considered as the abundance of cell types.
[0009] The specific technical solution of this invention is: a single-cell spatial transcriptome cell localization and type deconvolution algorithm based on feature decoupling, comprising the following steps:
[0010] The model requires paired single-cell gene transcription profiles and spatial transcriptomics data, obtained from public data sources.
[0011] We screened genes expressing all zeros, single-cell transcriptome data with all genes expressing zeros, and spatial transcriptome data, and then screened highly variable genes from paired single-cell transcriptome data and spatial transcriptome data.
[0012] The data were standardized, and the union of paired single-cell transcriptome data and the top 2000 highly variable genes from the spatial transcriptome data was retained.
[0013] Pairs of single-cell transcriptome data and spatial transcriptome data are input into the model. A common feature encoder decouples the corresponding shared features, while a domain-specific feature encoder decouples the corresponding domain-specific features. To achieve complete feature decoupling, orthogonal constraints are applied to the shared features and domain-specific features to ensure their independence.
[0014] Shared features from the single-cell domain are input into the cell type predictor, and a cross-entropy loss constraint is applied to both the predicted and actual cell types to ensure that the shared features contain cell type information. Shared features from the spatial transcriptome domain are input into the spatial location predictor, and a mean squared error loss constraint is applied to both the predicted and actual spatial coordinates to ensure that the shared features contain spatial location information.
[0015] Shared features and domain-specific features are concatenated, and the domain input features are reconstructed using a shared decoder. Reconstruction loss is used for constraints and overall training is performed to ensure that the domain-specific features retain the decoupled information.
[0016] Shared features and domain-specific features are concatenated, and the domain input features are reconstructed using a shared decoder. Reconstruction loss is used for constraints and overall training is performed to ensure that the domain-specific features retain the decoupled information.
[0017] The trained public encoder is used to extract shared features from the two domains. The shared features of the single-cell domain are input into the spatial location predictor to obtain the pseudo-spatial coordinates of the single cell. The shared features of the spatial transcriptome domain are input into the cell type predictor to obtain the cell type prediction score of the spatial transcriptome. This prediction score is regarded as the compositional abundance of the cell type.
[0018] The model network includes a public feature encoder that extracts shared features, two domain-specific feature encoders that extract domain-specific features, a shared decoder that reconstructs the domain input features, a cell type predictor that generates cell type prediction results, and a spatial location predictor that outputs spatial coordinates.
[0019] The present invention has the following beneficial effects:
[0020] 1. This feature transfer-based prediction algorithm for single-cell pseudospace and spatial transcriptomics cell type composition applies cross-domain transfer learning techniques to reveal cellular heterogeneity. For the first time, it integrates single-cell pseudospace reconstruction with cell type deconvolution of spatial transcriptomics probe points into a multi-task learning framework, achieving mutual supervision between the two modes and fully utilizing their complementary information. Through cross-domain transfer constraints, our model can effectively learn transferable shared information representations, thereby enhancing its generalization ability to unknown cell spaces and observed transcriptomics probe cell type composition. This approach overcomes the limitations of traditional models in revealing cellular heterogeneity, significantly improving the model's adaptability and accuracy.
[0021] 2. Our feature transfer-based prediction algorithm for single-cell pseudospace and spatial transcriptome cell type composition was comprehensively evaluated on various types of simulated and real biological datasets. Experimental results show that our model outperforms existing methods across all metrics, demonstrating higher prediction accuracy and robustness. Particularly in handling complex cancer datasets, our model exhibits superior performance, validating its potential and advantages in practical applications. Attached Figure Description
[0022] Figure 1 This is a diagram of the model algorithm framework;
[0023] Figure 2 Box plot of Pearson correlation coefficients comparing nine other algorithms on the same dataset.
[0024] Figure 3 A bar chart showing the Pearson correlation coefficients of different algorithms on the same dataset. Detailed Implementation
[0025] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.
[0026] Reference Figure 1 A single-cell spatial transcriptome cell localization and type deconvolution algorithm based on feature decoupling includes the following steps:
[0027] S1: Paired single-cell gene transcription profiles and spatial transcriptomics data obtained from public data sources. Genes expressing all zeros, single-cell transcription profiles with all genes expressing zero, and spatial transcriptomics data were screened. Highly variable genes were then selected from the paired single-cell and spatial transcriptomics data. The data were standardized, and the union of the top 2000 highly variable genes from each paired single-cell and spatial transcriptomics data was retained.
[0028] S2: Standardize the gene transcription profile data and divide the data into training, validation and test sets according to the drug-level data isolation strategy. Then, divide each dataset into multiple subsets according to cell line, and count the drug types in each subset. Divide the subset into two groups: group 1 and group 2 to ensure that the drug information of the two groups is completely different. Each group is then sorted according to drug name. Then, cyclic matching is performed on group 1 and group 2 to obtain multiple pairs of transcription profile data with the same cell but different drugs.
[0029] S3: Using a domain separation network, single-cell RNA sequencing is decoupled into two independent features. We use... This represents the single-cell RNA sequencing transcriptome, where N s This indicates the number of single-cell RNA sequencing samples. There are two encoders, a common encoder E. C (Shared parameters with the spatial transcriptomics domain public encoder) and domain-private encoder The expression profile is used as input, and two embeddings are output separately. and To ensure feature embedding decoupling, we require and They are mutually orthogonal, so that each embedding does not contain any information about the others. Therefore, we define the loss function as follows:
[0030]
[0031] At the same time, we require and The combination contains enough information to recover the original signal. In practical applications, both the single-cell RNA sequencing domain and the ST domain are reconstructed by the same decoder D with shared parameters. We will use two embeddings... and After splicing The input is fed into decoder D, and we require... as close as possible Therefore, we define the reconstruction loss of decoupled single-cell features as follows:
[0032]
[0033] Furthermore, we hope and Separated from each other, but The data should include information on cell type and spatial location, therefore we will Input to the single-cell domain predictor G s The shared encoder E obtains a score for single-cell type prediction, thus sharing weights. cIt can learn cell type information from the expression profile. The predictor G here... s It is implemented using a fully connected layer network, and we hope to predict the value. Compared with the true value To get as close as possible, we use cross-entropy loss to optimize the predictor:
[0034]
[0035] S4: We also used a domain separation network to decouple the spatial transcriptome into two independent features. We used Represents the spatial transcriptome, where N t This indicates the number of spatial transcriptome spots. There are two encoders, a common encoder E. C (Shared parameters with the single-cell domain public encoder) and private encoder The expression profile is used as input, and two embeddings are output separately. and To ensure feature embedding decoupling, we require and They are mutually orthogonal, so that each embedding does not contain any information about the others. Therefore, we define the loss function as follows:
[0036]
[0037] At the same time, we require and The combination of these two embeddings contains enough information to recover the original signal. and After splicing The input is fed into decoder D, and we require... as close as possible Therefore, we define the reconstruction loss of decoupled single-cell features as follows:
[0038]
[0039] Similarly, although we hope and They are separated from each other, but The data should include information on cell type and spatial location, therefore we will The predictor G input to the ST domain t This is used to obtain a predicted value for the spatial location of the spot, thus enabling the shared encoder E with shared weights. c It can learn spatial location information from the representation spectrum. The predictor G here... t Also implemented using a fully connected layer network, we hope to predict the value. Compared with the true value To get as close as possible, we use mean squared error loss to optimize the predictor:
[0040]
[0041] S6: The total loss function is defined as follows:
[0042]
[0043] The encoder / decoder network is implemented using a fully connected feedforward network with a ReLU activation function.
[0044] In the network model:
[0045] Domain-private feature encoder Extract single-cell domain private features from the input single-cell gene transcription profile.
[0046] Domain-private feature encoder Extract spatial transcriptome domain-specific features from the input spatial transcriptome data.
[0047] Common Feature Encoder E C : Extract shared features containing cell type and spatial location information from single-cell gene expression profile data and spatial transcriptome data.
[0048] Decoder D: Concatenates domain-private and shared features in the latent space and maps them back to the corresponding transcription profile.
[0049] Cell type predictor G s Predict cell type composition information from shared features.
[0050] Spatial Position Predictor G t Predicting spatial location information from shared features.
[0051] In this embodiment, paired single-cell transcriptome profiles and spatial transcriptomes from public data sources were used as experimental data. This dataset includes a mouse visual cortex dataset. The single-cell resolution spatial transcriptome data was gridded using a coarse-graining procedure introduced by SpatialDWLS, containing 189 probe points, each covering a 750×750 pixel region. To simulate gene expression profiles from probe points containing multiple cells, we aggregated the expression values of all cells within each probe point. Since both datasets are single-cell resolution spatial transcriptome data, the spatial location of individual cells is inherently determined, allowing us to evaluate the framework's performance in pseudo-space reconstruction tasks. We calculated the Pearson correlation coefficient between the predicted pseudo-space coordinates and the true location coordinates of single cells, as well as the Pearson correlation coefficient, structural similarity, and mean squared error between the predicted spatial transcriptome cell type composition and the true cell type composition, and compared these results with a benchmark model.
[0052] Experiments showed that the model improved single-cell pseudo-spatial prediction by 59% compared to the scSpace baseline model. It also outperformed Cell2location, DestVI, DSTG, RCTD, Seurat, Tangram, SPOTlight, Stereoscope, STRIDE, and SpatialDWLS in terms of Pearson correlation coefficient and structural similarity in spatial transcriptome cell type composition, with an average Pearson correlation coefficient of 0.79 and an average structural similarity of 0.69. All p-values were below 0.001. Regarding mean squared error, although some methods slightly outperformed our method, there was no statistically significant difference between the best method, Cell2location, and our method (mean mean squared error 0.15v8 0.16), with a p-value of 0.3.
[0053] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.
Claims
1. A single-cell spatial transcriptome cell localization and type deconvolution algorithm based on feature decoupling, characterized in that: Including the following steps: 1) Data screening and preprocessing: Screening for genes expressing all zeros, single-cell transcriptome data with all genes expressing zeros, and spatial transcriptome data, and screening for highly variable genes from paired single-cell transcriptome data and spatial transcriptome data. 2) The data were standardized and the union of paired single-cell transcriptome data and the top 2000 highly variable genes in the spatial transcriptome data was retained. 3) Paired single-cell transcriptome data and spatial transcriptome data are input into the model. The common feature encoder decouples the corresponding shared features, and the domain-specific feature encoder decouples the corresponding domain-specific features. To achieve complete feature decoupling, orthogonal constraints are applied to the shared features and domain-specific features to ensure their independence. 4) Input the shared features of the single-cell domain into the cell type predictor, and apply cross-entropy loss constraints to the predicted cell types and the actual cell types to ensure that the shared features contain cell type information. Input the shared features of the spatial transcriptome domain into the spatial location predictor, and apply mean squared error loss constraints to the predicted spatial coordinates and the actual spatial coordinates to ensure that the shared features contain spatial location information. 5) Concatenate the shared features and domain-specific features, reconstruct the domain input features (i.e., the gene expression features generated in step 2) using the shared decoder, constrain them with reconstruction loss and train the whole system to ensure that the domain-specific features retain the decoupled information. 6) Use the trained common encoder to extract the shared features of the two domains, and input the shared features of the single cell domain into the spatial location predictor to obtain the pseudo spatial coordinates of the single cell. Input the shared features of the spatial transcriptome domain into the cell type predictor to obtain the cell type prediction score of the spatial transcriptome. The prediction score is regarded as the composition abundance of the cell type. In step 1), single-cell gene transcription profile data and spatial transcriptome data are obtained from public data sources. In step 2), the paired single-cell gene transcription data and spatial transcriptome data are derived from the same experiment or sample.
2. The single-cell spatial transcriptome cell localization and type deconvolution algorithm based on feature decoupling according to claim 1, characterized in that: Highly variable genes were screened from paired single-cell gene transcription data and spatial transcriptome data, and the union of the top 2000 genes was retained.
3. The single-cell spatial transcriptome cell localization and type deconvolution algorithm based on feature decoupling according to claim 1, characterized in that: In step 3), the public feature encoder and the domain-specific feature encoder are fully connected neural networks responsible for decoupling and separating shared features and domain-specific features from the transcriptional spectrum data, and mapping the shared features to a low-dimensional latent space.
4. The single-cell spatial transcriptome cell localization and type deconvolution algorithm based on feature decoupling according to claim 1, characterized in that: In step 4), the common feature encoder and predictor work together to learn the shared information of the shared features. Through the intervention of the predictor, the information of cell type and spatial location is preserved in the shared features.
5. The single-cell spatial transcriptome cell localization and type deconvolution algorithm based on feature decoupling according to claim 1, characterized in that: In step 5), the domain-private encoder, public encoder, and public decoder work together to perform feature decoupling tasks. Through the reconstruction task of the decoder, the ground state information of domain-specific features is effectively preserved.
6. The single-cell spatial transcriptome cell localization and type deconvolution algorithm based on feature decoupling according to claim 1, characterized in that: After the predictor processes the shared features in the latent space, it can accurately predict and output the pseudo-spatial coordinates of a single cell and the cell type composition abundance of the spatial transcriptome.