Method and device for constructing deconvolution model, terminal and medium

Abstract

The invention provides a deconvolution model construction method and device, a terminal and a medium, and the method comprises the steps: generating simulated space transcriptome data with a cell type proportion label based on single cell transcriptome data, and obtaining real space transcriptome data obtained through sampling based on a 3D concentric sphere sampling strategy, and based on a comparative learning strategy and a domain adversarial network, training an encoder and an initial prediction model by using target space transcriptome data containing differential expression genes obtained by screening from simulation data and real data, and obtaining a feature extraction model for extracting embedded information and a prediction model for predicting a cell type proportion. According to the method, deep learning is combined, target space transcriptome data is used as encoder input, encoder output is used as prediction model input, the model is trained based on domain adversarial learning and contrast learning strategies, a model with good generalization ability is obtained, and high-precision deconvolution and cell type proportion prediction of 3D space transcriptome data are realized.

Description

Technical Field

[0001] The present invention relates to the field of bioinformatics technology, and particularly to a method, device, terminal and medium for constructing a deconvolution model. Background Art

[0002] Spatial transcriptome is a very important type of data in life science research. It not only reveals the spatial distribution of gene expression in tissues, but also plays a key role in studying the structure and function of complex biological systems. In the integrated analysis of spatial transcriptome and single-cell transcriptome, how to improve the resolution of the spatial transcriptome and analyze the proportion of cell types is a challenging task, and there are mainly the following core challenges:

[0003] First, the spatial transcriptome data itself has characteristics such as low resolution, strong spatial dependence, and high data heterogeneity. Even for high-precision technologies, such as MERFISH (Multiplexed Error-Robust Fluorescence In Situ Hybridization), the sampling sites are close to the cell level, but it is still impossible to accurately identify cell boundaries, resulting in large errors in inferring the proportion of cell types. 3D (Three-Dimensional) spatial transcriptome data needs to solve the cross-modal alignment problem between cryosection images of samples and transcriptome information, especially the matching and integration of edge information of sections. Existing methods are difficult to effectively handle such complex scenarios.

[0004] Second, each site in the spatial transcriptome data is usually composed of multiple cell types. For high-precision spatial transcriptomes, there are also cases where the center of the sequencing site is at the junction of cells, making its deconvolution task a complex problem of multi-label and multi-classification.

[0005] Third, single-cell RNA (Ribonucleic Acid) sequencing (scRNA-seq) makes it possible to perform transcriptome analysis at the single-cell level. However, due to the lack of 3D spatial information, its application is limited, which hinders the accurate interpretation of cell-cell relationships and affects the construction of tissue models. Tissue models are crucial for downstream applications, such as the design of cell factories. Although the third-generation spatial transcriptomics solves the problem of lack of 3D spatial information by retaining spatial information during sequencing, no single spatial transcriptome technology can meet all application scenarios. Moreover, currently commonly used technologies, such as Slide-seq, 10x Genomics Visium, etc., still face limitations in spatial resolution, resulting in the inability to accurately capture cell positions and boundaries, and further affecting the restoration of real cell boundaries.

[0006] Fourth, although both traditional machine learning-based and deep learning-based deconvolution can achieve high-precision deconvolution to a certain extent, there are still limitations in the utilization of spatial information. For example, in traditional machine learning-based deconvolution, such as traditional log-linear models and Bayesian models based on negative binomial distributions, they perform poorly in spatial context modeling. In deep learning model-based deconvolution, such as DeepST and CellDART, although they perform well in spatial domain recognition, their spatial interpretability is poor. Another example is that although GraphST optimizes cell position inference, its spatial interpretability still lags behind traditional machine learning models.

[0007] Fifth, there are significant differences in spatial transcriptome data of different tissue samples. However, most models lack adaptability to tissue specificity, which limits their application in multiple scenarios and results in insufficient cross-tissue generalization ability.

[0008] In summary, the existing methods have deficiencies in embedding spatial information or capturing complex spatial relationships between cells. Traditional machine learning methods often perform poorly in representing spatial context, resulting in the need to optimize deconvolution results. In addition, many methods rely on simplified models and fail to fully utilize rich spatial and histological data, thus having limitations in terms of accuracy and interpretability.

[0009] Therefore, it has become an urgent concern for those skilled in the art to provide a model that can accurately infer cell type proportions, fully utilize the neighborhood relationships between spatial sites, and also has good generalization ability to cope with tissue specificity to achieve high-precision deconvolution of 3D spatial transcriptome data and prediction of cell type proportions. Summary of the Invention

[0010] The technical problem to be solved by the present invention is to provide a method, device, terminal and medium for constructing a deconvolution model, which can achieve high-precision deconvolution of 3D spatial transcriptome data and prediction of cell type proportions in view of the above-mentioned defects of the prior art.

[0011] The technical solution adopted by the present invention to solve the technical problem is as follows:

[0012] A method for constructing a deconvolution model, wherein the method includes:

[0013] Generating simulated spatial transcriptome data with cell type proportion labels based on single-cell transcriptome data of the same tissue;

[0014] Obtaining real spatial transcriptome data sampled based on a 3D concentric sphere sampling strategy;

[0015] Perform differential expression gene screening based on the simulated spatial transcriptome data and the real spatial transcriptome data to obtain target spatial transcriptome data containing differentially expressed genes;

[0016] Based on the contrastive learning strategy and the domain adversarial network, and use the target spatial transcriptome data to train a pre-constructed encoder and an initial prediction model, to obtain a trained feature extraction model for extracting embedding information and a prediction model for predicting the cell type proportion; wherein, during the training process, the embedding information output by the encoder is used as the input of the initial prediction model and the domain adversarial network.

[0017] In one implementation, the generating of the simulated spatial transcriptome data with cell type proportion labels based on the single-cell transcriptome data of the same tissue includes:

[0018] Generate the true cell type proportion labels of the single-cell transcriptome data of the same tissue, and generate the simulated spatial transcriptome data with cell type proportion labels based on the true cell type proportion labels and the single-cell transcriptome data of the same tissue.

[0019] In one implementation, the performing of differential expression gene screening based on the simulated spatial transcriptome data and the real spatial transcriptome data to obtain target spatial transcriptome data containing differentially expressed genes includes:

[0020] Use the t-test method to perform differential analysis on each gene in the simulated spatial transcriptome data and the real spatial transcriptome data to calculate the corresponding p-value;

[0021] Judge whether the p-value is less than a preset p-value threshold to obtain a corresponding judgment result, and perform differential expression gene screening based on the judgment result to obtain target spatial transcriptome data containing differentially expressed genes.

[0022] In one implementation, the performing of differential expression gene screening based on the judgment result to obtain target spatial transcriptome data containing differentially expressed genes includes:

[0023] When the judgment result indicates that the p-value is less than the preset p-value threshold, screen the differentially expressed genes corresponding to the p-value in the simulated spatial transcriptome data and the real spatial transcriptome data to obtain target spatial transcriptome data containing differentially expressed genes.

[0024] In one implementation, the preset p-value threshold is 0.005.

[0025] In one implementation, based on the contrastive learning strategy and the domain adversarial network, and using the target space transcriptome data to train a pre-constructed encoder and an initial prediction model, to obtain a trained feature extraction model for extracting embedding information and a prediction model for predicting the cell type ratio, including:

[0026] Input the target space transcriptome data into the pre-constructed encoder for training on extracting embedding information;

[0027] Based on the contrastive learning strategy, optimize the spatial information in the embedding information to determine the mutual relationship between spatial sites, and calculate the corresponding contrastive loss;

[0028] Input the embedding information into the domain adversarial network for domain adversarial training to calculate the corresponding binary cross-entropy loss, and input the embedding information into the initial prediction model for prediction training of the cell type ratio, and calculate the corresponding L1 loss;

[0029] Until the contrastive loss, the binary cross-entropy loss, and the L1 loss meet the preset loss conditions, obtain the trained feature extraction model for extracting the embedding information and the prediction model for predicting the cell type ratio.

[0030] In one implementation, during the process of predicting the cell type ratio by the prediction model, it further includes:

[0031] Process the high-dimensional embedding representation generated by the feature extraction model according to the preset neural network operations to generate the probability distribution of the corresponding cell types, and predict the cell type ratio based on the probability distribution of the cell types; wherein, the preset neural network operations include: linear transformation operation, Leaky ReLU activation function, Dropout operation, and softmax function.

[0032] The present invention also discloses a deconvolution model construction device, wherein, the device includes:

[0033] A data simulation module, configured to generate simulated space transcriptome data with cell type ratio labels based on the same tissue single-cell transcriptome data;

[0034] A data sampling module, configured to obtain real space transcriptome data sampled based on the 3D concentric sphere sampling strategy;

[0035] A gene screening module, configured to screen for differentially expressed genes based on the simulated space transcriptome data and the real space transcriptome data to obtain target space transcriptome data containing differentially expressed genes;

[0036] A model training module, which is used to train a pre-constructed encoder and an initial prediction model based on a contrastive learning strategy and a domain adversarial network, and by using the target space transcriptome data, to obtain a trained feature extraction model for extracting embedding information and a prediction model for predicting the proportion of cell types; wherein, during the training process, the embedding information output by the encoder is used as the input of the initial prediction model and the domain adversarial network.

[0037] The present invention also discloses a terminal, which includes: a memory, a processor, and a deconvolution model construction program stored on the memory and executable on the processor, and when the deconvolution model construction program is executed by the processor, it implements the steps of the deconvolution model construction method as described above.

[0038] The present invention also discloses a computer-readable storage medium, wherein the computer-readable storage medium stores a computer program, and the computer program can be executed to implement the steps of the deconvolution model construction method as described above.

[0039] A method, device, terminal, and medium for constructing a deconvolution model provided by the present invention. The method for constructing the deconvolution model includes: generating simulated spatial transcriptome data with cell type proportion labels based on single-cell transcriptome data of the same tissue; obtaining real spatial transcriptome data sampled based on a 3D concentric sphere sampling strategy; screening for differentially expressed genes based on the simulated spatial transcriptome data and the real spatial transcriptome data to obtain target spatial transcriptome data containing differentially expressed genes; training a pre-constructed encoder and an initial prediction model based on a contrast learning strategy and a domain adversarial network using the target spatial transcriptome data to obtain a trained feature extraction model for extracting embedding information and a prediction model for predicting cell type proportions; wherein, during the training process, the embedding information output by the encoder serves as the input to the initial prediction model and the domain adversarial network. It can be seen from this that the present invention generates simulated spatial transcriptome data based on single-cell transcriptome data of the same tissue, obtains real spatial transcriptome data sampled based on a 3D concentric sphere sampling strategy, then screens for differentially expressed genes based on the simulated spatial transcriptome data and the real spatial transcriptome data, and further, based on a contrast learning strategy and a domain adversarial network, inputs the obtained target spatial transcriptome data containing differentially expressed genes into the pre-constructed encoder for training, and uses the output of the encoder as the input to the initial prediction model for training to obtain a trained feature extraction model for extracting embedding information and a prediction model for predicting cell type proportions. That is, the present application combines deep learning techniques, uses the target spatial transcriptome data containing differentially expressed genes as the input to the encoder, and the output of the encoder as the input to the prediction model, so as to train the encoder and the prediction model based on domain adversarial learning and contrast learning strategies, thereby being able to accurately infer cell type proportions, and having good generalization ability to handle tissue specificity, and a feature extraction model and a prediction model for high-precision deconvolution of 3D spatial transcriptome data and prediction of cell type proportions are realized. Description of the Drawings

[0040] Figure 1 is a flowchart of a preferred embodiment of the method for constructing a deconvolution model in the present invention;

[0041] Figure 2 is a schematic diagram of training a spatial transcriptome data deconvolution model based on spatial information and a domain adversarial network disclosed by the present invention;

[0042] Figure 3 is a schematic diagram of comparing a specific 2D deconvolution method for spatial transcriptome data in the present invention;

[0043] Figure 4 is a schematic diagram of a specific contrast learning 2D circular sampling in the present invention;

[0044] Figure 5 It is a functional principle block diagram of a preferred embodiment of the deconvolution model construction device in the present invention;

[0045] Figure 6 It is a functional principle block diagram of a preferred embodiment of the terminal in the present invention. Specific embodiments

[0046] To make the objectives, technical solutions and advantages of the present invention clearer and more definite, the following further elaborates on the present invention with reference to the accompanying drawings and by way of examples. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0047] Please refer to Figure 1 , Figure 1 It is a flowchart of the deconvolution model construction method in the present invention. As Figure 1 shown, the deconvolution model construction method described in the embodiments of the present invention includes:

[0048] Step S11: Generate simulated spatial transcriptome data with cell type proportion labels based on single-cell transcriptome data of the same tissue.

[0049] In this embodiment, based on the single-cell transcriptome data of the same tissue, simulated spatial transcriptome data is generated, and the simulated spatial transcriptome data is spatial transcriptome data with cell type proportion labels.

[0050] Specifically, generate true cell type proportion labels for the single-cell transcriptome data of the same tissue, and generate simulated spatial transcriptome data with cell type proportion labels based on the true cell type proportion labels and the single-cell transcriptome data of the same tissue. It can be understood that accurate proportion labels are generated, and then these proportion labels are combined with the single-cell transcriptome data to generate simulated spatial transcriptome data with cell type proportion labels.

[0051] In this embodiment, after generating simulated spatial transcriptome data with cell type proportion labels based on the single-cell transcriptome data of the same tissue, it may further specifically include: allocating corresponding position information labels according to the similarity of gene expression at each locus in the simulated spatial transcriptome data. For example, using the k-means clustering algorithm to add coordinate information to the spatial transcriptome data, so as to be able to construct a data structure that better conforms to the spatial characteristics. That is to say, based on the single-cell transcriptome data of the same tissue, simulated spatial transcriptome data is generated, and the spatial coordinates are optimized by combining the k-means clustering.

[0052] Step S12: Obtain true spatial transcriptome data sampled based on the 3D concentric sphere sampling strategy.

[0053] It is understandable that real - space transcriptome data is sampled from different types of samples based on a 3D concentric sphere sampling strategy. It should be noted that after obtaining the real - space transcriptome data, a clarity - reduction process is performed on it to obtain simulated low - clarity space transcriptome data. Moreover, based on the high - clarity real - space transcriptome data, low - resolution locus data is generated, so that the model can attempt to identify the proportion of cell types and restore the spatial transcriptome data of the single - cell type hierarchy.

[0054] Step S13: Based on the simulated space transcriptome data and the real - space transcriptome data, differential expression gene screening is performed to obtain target space transcriptome data containing differentially expressed genes.

[0055] In this embodiment, for the simulated space transcriptome data and the real - space transcriptome data, differential expression gene screening is performed. Specifically, the t - test method is used to perform differential analysis on each gene in the simulated space transcriptome data and the real - space transcriptome data to calculate the corresponding p - value; it is determined whether the p - value is less than a preset p - value threshold to obtain the corresponding judgment result, and based on the judgment result, differential expression gene screening is performed to obtain target space transcriptome data containing differentially expressed genes. When the judgment result indicates that the p - value is less than the preset p - value threshold, the differentially expressed genes corresponding to the p - values of the simulated space transcriptome data and the real - space transcriptome data are screened to obtain target space transcriptome data containing differentially expressed genes. Among them, the preset p - value threshold can be 0.005. It is understandable that differentially expressed genes with p - values less than 0.005 are screened to reduce the dimension and enhance the quality of training data, so as to be normalized with the space transcriptome data with reduced clarity, and then can be used for domain adversarial learning to improve the deconvolution accuracy and model generalization ability.

[0056] For example, to solve the problems that many existing space transcriptome data have low resolution, the cell boundaries of high - resolution data are not clear, and single - cell transcriptome data lacks spatial location information in real - space transcriptome data, based on a randomized simulated data generation method, simulated space transcriptome data with cell - type proportion labels can be generated. After obtaining the real - space transcriptome data sampled based on the 3D concentric sphere sampling strategy, a resolution - reduction process can be performed on it to obtain simulated low - resolution space transcriptome data with spatial information. Then, in data processing, the real - space transcriptome data and the simulated space transcriptome data can be integrated by normalization means, and then key genes (differentially expressed genes) with p - values less than 0.005 are screened through the t - test method to reduce the data dimension and improve the training efficiency. Among them, the combination of the real - space transcriptome data and the simulated space transcriptome data can provide a basis for domain adversarial learning, effectively enhancing the application of the model in cell - boundary recognition, supporting more accurate spatial transcriptome deconvolution cell - type classification, and the division of cell boundaries.

[0057] Step S14: Based on the contrastive learning strategy and the domain adversarial network, and using the target space transcriptome data, train the pre-constructed encoder and the initial prediction model to obtain a trained feature extraction model for extracting embedding information and a prediction model for predicting the proportion of cell types; wherein, during the training process, the embedding information output by the encoder is used as the input of the initial prediction model and the domain adversarial network.

[0058] In this embodiment, the target space transcriptome data containing differentially expressed genes is used as the input of the encoder, and the output of the encoder is used as the input of the initial prediction model. Thus, based on the domain adversarial learning and contrastive learning strategies, the encoder and the prediction model are trained, so as to obtain a feature extraction model and a prediction model that can accurately infer the proportion of cell types and have good generalization ability to cope with tissue specificity, in order to achieve high-precision deconvolution of 3D space transcriptome data and prediction of the proportion of cell types.

[0059] Specifically, input the target space transcriptome data containing differentially expressed genes into the pre-constructed encoder for training the extraction of embedding information; optimize the spatial information in the embedding information based on the contrastive learning strategy to determine the mutual relationship between spatial sites, and calculate the corresponding contrast loss; input the embedding information into the domain adversarial network for domain adversarial training to calculate the corresponding binary cross-entropy loss; and input the embedding information into the initial prediction model for training the prediction of the proportion of cell types, and calculate the corresponding L1 loss, until the contrast loss, the binary cross-entropy loss, and the L1 loss meet the preset loss conditions, to obtain a trained feature extraction model for extracting embedding information and a prediction model for predicting the proportion of cell types.

[0060] Among them, during the training process, contrastive learning is used to extract the mutual relationships of spatial points, and the generalization ability of the model to real data is improved through a Domain-Adversarial Network (DAN). Finally, a feature extraction model with high generality is trained. And during the process of predicting the cell type ratio through the prediction model, the prediction model processes the high-dimensional embedding representation generated by the feature extraction model according to the preset neural network operations, generates the probability distribution of the corresponding cell type, and predicts the cell type ratio based on the probability distribution of the cell type. Among them, the preset neural network operations include: linear transformation operation, Leaky ReLU activation function, Dropout operation, and softmax function. That is to say, the prediction model processes the high-dimensional embedding information generated by the encoder through the linear transformation operation, Leaky ReLU activation function, Dropout operation, and softmax function to achieve accurate prediction of the cell type ratio, and the overall model performance can be improved through optimization strategies. Specifically, the high-dimensional embedding representation generated by the feature extraction model is dimensionally reduced through linear transformation, and the Leaky ReLU activation function is used to introduce non-linearity, thereby enhancing the expression ability for complex features. Subsequently, the Dropout operation is used to reduce the overfitting risk and improve the robustness of the model. In the final output stage, the embedding representation undergoes feature transformation through another linear layer, and the softmax function is used to generate the probability distribution of the cell type to complete the prediction of the cell type ratio. Among them, to optimize the performance of the prediction model, the Adam optimizer can be used to jointly train the parameters of the encoder and the prediction model, effectively improving the prediction accuracy of the model, and the decoder module (i.e., the predict part of the model) generates spatial transcriptomic data at the single-cell level.

[0061] It should be noted that the pre-constructed encoder is used to effectively represent the gene expression information in the target spatial transcriptomic data, combines the gene expression data of each spatial site with its corresponding two-dimensional spatial coordinates to generate high-dimensional embedding vectors. That is to say, in a way of gradually reducing dimensions, the encoder removes redundant features while retaining key information, providing a simplified and efficient representation for the subsequent learning process. That is to say, through the encoder structure, the gene expression in the target spatial transcriptomic data is effectively represented, and the embedding information of the gene expression is extracted. Subsequently, using the contrastive learning strategy, the spatial information in the embedding information is optimized to capture the mutual relationships of spatial points. Finally, through the domain adversarial network, the simulated spatial transcriptomic data and the real spatial transcriptomic data can be adapted, so as to be able to construct a feature extraction model with high generality, thereby improving the accuracy and applicability of spatial transcriptomic deconvolution.

[0062] Moreover, after extracting the embedding information through the encoder, a contrastive learning strategy is introduced to optimize the spatial information representation in the embedding information. In contrastive learning, by constructing positive and negative sample pairs, the embedding distances between adjacent points in the three-dimensional space are shortened, while the embedding distances between non-adjacent points in the space are lengthened, so as to better capture the mutual relationships between spatial points. Among them, contrastive learning enhances the model's ability to capture and utilize spatial relationships by distinguishing positive and negative sample pairs based on the 3D spatial proximity relationship. And during the sampling process of real-space transcriptome data, by adopting a 3D concentric sphere sampling strategy to dynamically define positive and negative sample pairs, the complexity of tissue morphological structures can be addressed, such as the annular hierarchical structure of human embryos at the cs8 stage, further enhancing the model's sensitivity to spatial context information. Among them, by establishing a deconvolution framework through spherical sampling and contrastive learning, 3D spatial information can be better integrated.

[0063] It should also be noted that to address the domain difference problem between simulated spatial transcriptome data and real-space transcriptome data, a Domain-Adversarial Network (DAN) is adopted. Then, through adversarial training, the domain adversarial network aligns the two data distributions, reducing the embedding differences between simulated spatial transcriptome data and real-space transcriptome data, and the similarity of different data distributions can be quantified and optimized through the Maximum Mean Discrepancy (MMD) loss function. That is to say, in the technical solution of this application, the combination of domain adversarial technology significantly improves the model's ability to identify the cell boundaries of real data. That is, by introducing a domain adversarial network during the training process, the noise generated by simulated spatial data with cell type labels can be reduced, thereby improving the overall generalization ability of the deconvolution model.

[0064] It can be seen that in the embodiments of the present invention, based on the single-cell transcriptome data of the same tissue, simulated spatial transcriptome data is generated, and real spatial transcriptome data obtained by sampling based on the 3D concentric sphere sampling strategy is acquired. Then, differential expression gene screening is performed based on the simulated spatial transcriptome data and the real spatial transcriptome data. Furthermore, based on the contrast learning strategy and the domain adversarial network, the target spatial transcriptome data containing differentially expressed genes obtained by screening is input into a pre-constructed encoder for training, and the output of the encoder is used as the input of the initial prediction model for training, obtaining a trained feature extraction model for extracting embedding information and a prediction model for predicting the cell type ratio. That is, in this application, by combining deep learning techniques, the target spatial transcriptome data containing differentially expressed genes is used as the input of the encoder, and the output of the encoder is used as the input of the prediction model, so as to train the encoder and the prediction model based on the domain adversarial learning and contrast learning strategies, thereby being able to accurately infer the cell type ratio and having good generalization ability to cope with tissue specificity, and realizing high-precision deconvolution of 3D spatial transcriptome data and prediction of cell type ratio.

[0065] For example, as shown in Figure 2 Based on the deep learning-based deconvolution model, namely ST-deconv3d (a spatial transcriptome data deconvolution model based on spatial information and domain adversarial network), during data preprocessing, accurate ratio labels are generated, and then by combining these ratio labels with single-cell transcriptome data, simulated spatial transcriptome data with cell type ratio labels is generated. According to the similarity of gene expression at each locus in the simulated spatial transcriptome data, corresponding position information labels are assigned, and sampling is performed from different types of samples based on the 3D concentric sphere sampling strategy to obtain the sampled real spatial transcriptome data. Then, differential expression gene screening is performed on the simulated spatial transcriptome data and the real spatial transcriptome data by means of t-test, and the target spatial transcriptome data containing differentially expressed genes obtained by screening is aligned for further analysis. The preprocessed data (i.e., the target spatial transcriptome data containing differentially expressed genes obtained by screening) is input into an encoder (Encoder) to extract features. During the training process, there are a contrast loss (CLloss) of a contrast learning module (CL module), an L1 loss (L1loss) of the prediction model, and a binary cross-entropy loss (BCEloss) of a domain classification module (domain adversarial network). In the evaluation mode, the encoder extracts features and inputs the features into the domain classification module to calculate the binary cross-entropy loss (BCEloss). That is to say, after t-test, the screened differentially expressed gene data is input into the encoder to extract features, and the binary cross-entropy loss is calculated in the domain classification module to promote fine analysis and verification.

[0066] It should be noted that ST-deconv-3d utilizes spatial transcriptome data and single-cell RNA sequencing (scRNAseq) data to effectively improve the resolution of low-resolution spatial transcriptomes, and has made significant progress in predicting cell type distributions. Moreover, it integrates more spatial relationships to accurately classify cell types and can be extended to infer cell interactions in multiple biological contexts. In the core architecture of ST-deconv-3d, it combines a contrastive learning strategy and the generalization ability of a Domain-Adversarial Network (DAN). Through contrastive learning, it can enhance the similarity in the hidden layer embeddings related to the true distances between spatial loci, thereby effectively introducing spatial information and significantly improving the ability to capture the relationships between spatial loci. Additionally, by introducing spatial locus information from spatial transcriptome data through contrastive learning, it enhances the ability to analyze cell-cell relationships. Moreover, ST-deconv-3d can autonomously generate spatial transcriptome data for training, which provides a new approach to improving the resolution of spatial transcriptome data. That is, ST-deconv-3d can achieve high-precision deconvolution of spatial transcriptome data and contribute to understanding cell interactions in complex tissue structures. For currently emerging spatial transcriptome data with higher resolution and continuous sections, it can identify cell boundaries, utilize more comprehensive 3D spatial information to restore cells, and thus achieve the extension from spatiality to the 3D level.

[0067] After training is completed, the model is used to predict the cell type proportions. Simulated spatial transcriptome data can be used for testing to evaluate the prediction ability of the model. For example, on CARD (Comprehensive Adaptive Randomized Data), the dataset can be divided into a training set and a test set in a ratio of 8:2 and compared with other methods in a comparative experiment. See Figure 3 As shown, compared with the 2D deconvolution method for the simulated spatial transcriptome data of the mouse olfactory bulb (MOB) generated by CARD, that is, comparing the Root Mean Square Error (RMSE) of different deconvolution methods on spatial transcriptome data with different levels of spatial correlation, the results show that the two-dimensional ST-deconv method performs better in deconvolution performance. Since the present invention has achieved good results in the deconvolution task of two-dimensional spatial transcriptome data, theoretically, if more effective 3D spatial information is added, the spatial organizational structure of cells can be modeled more accurately, thereby further improving the accuracy of deconvolution and the fidelity of spatial features. Among them, 3D spatial information can better capture the interactions between cells at different levels, provide richer spatial information for the model, and is expected to further optimize the deconvolution results.

[0068] And, referring to Figure 4 As shown, for the mouse olfactory bulb data, the threshold is defined according to the size of the regional center site, and the radius of the positive sample site is established accordingly. For example, the data within the range of 2 to 3 sites from the center site is excluded from the sampling process, while the range of 3 to 5 sites is defined as the range of the negative sample site.

[0069] In one embodiment, as Figure 5 shown, based on the above deconvolution model construction method, the present invention also correspondingly provides a deconvolution model construction device, including:

[0070] A data simulation module 11, configured to generate simulated spatial transcriptome data with cell type proportion labels based on the same tissue single-cell transcriptome data;

[0071] A data sampling module 12, configured to obtain real spatial transcriptome data sampled based on a 3D concentric sphere sampling strategy;

[0072] A gene screening module 13, configured to screen for differentially expressed genes based on the simulated spatial transcriptome data and the real spatial transcriptome data to obtain target spatial transcriptome data containing differentially expressed genes;

[0073] A model training module 14, configured to train a pre-constructed encoder and an initial prediction model based on a contrastive learning strategy and a domain adversarial network, and use the target spatial transcriptome data to obtain a trained feature extraction model for extracting embedding information and a prediction model for predicting cell type proportions; wherein, during the training process, the embedding information output by the encoder serves as the input to the initial prediction model and the domain adversarial network.

[0074] Figure 6 This is a schematic structural diagram of the terminal provided in the embodiments of the present application. The terminal may include:

[0075] A memory 501, a processor 502, and a computer program stored on the memory 501 and executable on the processor 502.

[0076] When the processor 502 executes the program, it implements the deconvolution model construction method provided in the above embodiments.

[0077] Further, the terminal further includes:

[0078] A communication interface 503, configured for communication between the memory 501 and the processor 502.

[0079] The memory 501 is used to store a computer program executable on the processor 502.

[0080] The memory 501 may include high-speed RAM memory and may also include non-volatile memory, such as at least one disk memory.

[0081] If the memory 501, the processor 502, and the communication interface 503 are implemented independently, the communication interface 503, the memory 501, and the processor 502 can be interconnected via a bus and communicate with each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, an Extended Industry Standard Architecture (EISA) bus, or the like. The bus can be divided into an address bus, a data bus, a control bus, etc. For ease of representation, only one line is shown in the figure, but it does not mean that there is only one bus or one type of bus.

[0082] Optionally, in a specific implementation, if the memory 501, the processor 502, and the communication interface 503 are integrated on a single chip, the memory 501, the processor 502, and the communication interface 503 can communicate with each other via an internal interface.

[0083] The processor 502 may be a Central Processing Unit (CPU), or an Application Specific Integrated Circuit (ASIC), or one or more integrated circuits configured to implement the embodiments of the present application.

[0084] This embodiment also provides a computer-readable storage medium, on which a computer program is stored, and when the program is executed by a processor, the above-mentioned deconvolution model construction method is implemented.

[0085] Those skilled in the art will readily conceive of other embodiments of the present invention after considering the specification and practicing the invention disclosed herein. The present invention is intended to cover any variations, uses, or adaptations of the present invention, which follow the general principles of the present invention and include known common knowledge or conventional technical means in the technical field not disclosed in the present application. The specification and the embodiments are only regarded as exemplary, and the true scope and spirit of the present invention are pointed out by the claims.

[0086] In the description of this specification, the descriptions referring to terms such as "one embodiment", "some embodiments", "example", "specific example", or "some examples" etc. mean that the specific features, structures, materials, or characteristics described in connection with the embodiment or example are included in at least one embodiment or example of this application. In this specification, the schematic expressions of the above terms do not necessarily refer to the same embodiment or example. Moreover, the specific features, structures, materials, or characteristics described can be combined in a suitable manner in any one or N embodiments or examples. In addition, without contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of different embodiments or examples.

[0087] The logic and / or steps represented in the flowchart or described in other ways herein, for example, can be considered as a definite sequence list of executable instructions for implementing logical functions, and can be specifically implemented in any computer-readable medium for use by an instruction execution system, apparatus, or device (such as a computer-based system, a system including a processor, or other systems that can read and execute instructions from the instruction execution system, apparatus, or device), or in combination with these instruction execution systems, apparatuses, or devices.

[0088] It should be understood that each part of this application can be implemented by hardware, software, firmware, or a combination thereof. In the above embodiment, the N steps or methods can be implemented by software or firmware stored in a memory and executed by a suitable instruction execution system. If implemented by hardware, as in another embodiment, any one or a combination of the following technologies well-known in the art can be used: discrete logic circuits with logic gate circuits for implementing logical functions on data signals, application-specific integrated circuits with suitable combinational logic gate circuits, programmable gate arrays (PGA, Programmable Gate Array), field-programmable gate arrays (FPGA, Field-Programmable Gate Array), etc.

[0089] It should be understood that the application of the present invention is not limited to the above examples. For those of ordinary skill in the art, improvements or transformations can be made according to the above description, and all such improvements and transformations should fall within the protection scope of the appended claims of the present invention.

Claims

1. A method for constructing a deconvolution model, characterized in that The method includes: Generating simulated spatial transcriptome data with cell type proportion labels based on single-cell transcriptome data of the same tissue; Obtaining real spatial transcriptome data sampled based on a 3D concentric sphere sampling strategy; Screening for differentially expressed genes based on the simulated spatial transcriptome data and the real spatial transcriptome data to obtain target spatial transcriptome data containing differentially expressed genes; Training a pre-constructed encoder and an initial prediction model based on a contrast learning strategy and a domain adversarial network, and using the target spatial transcriptome data to obtain a trained feature extraction model for extracting embedding information and a prediction model for predicting cell type proportions; wherein, during the training process, the embedding information output by the encoder is used as the input of the initial prediction model and the domain adversarial network.

2. The method for constructing a deconvolution model according to claim 1, wherein The generating of simulated spatial transcriptome data with cell type proportion labels based on single-cell transcriptome data of the same tissue includes: Generating real cell type proportion labels for the single-cell transcriptome data of the same tissue, and generating simulated spatial transcriptome data with cell type proportion labels based on the real cell type proportion labels and the single-cell transcriptome data of the same tissue.

3. The method for constructing a deconvolution model according to claim 1, wherein The screening for differentially expressed genes based on the simulated spatial transcriptome data and the real spatial transcriptome data to obtain target spatial transcriptome data containing differentially expressed genes includes: Performing differential analysis on each gene in the simulated spatial transcriptome data and the real spatial transcriptome data using a t-test method to calculate the corresponding p-value; Judging whether the p-value is less than a preset p-value threshold to obtain a corresponding judgment result, and screening for differentially expressed genes based on the judgment result to obtain target spatial transcriptome data containing differentially expressed genes.

4. The method for constructing a deconvolution model according to claim 3, wherein The screening for differentially expressed genes based on the judgment result to obtain target spatial transcriptome data containing differentially expressed genes includes: When the judgment result indicates that the p-value is less than the preset p-value threshold, screening the differentially expressed genes corresponding to the p-value in the simulated spatial transcriptome data and the real spatial transcriptome data to obtain target spatial transcriptome data containing differentially expressed genes.

5. The method for constructing a deconvolution model according to claim 4, wherein The preset p-value threshold is 0.

005.

6. The method for constructing a deconvolution model according to any one of claims 1 to 5, characterized in that The training of a pre-constructed encoder and an initial prediction model based on a contrast learning strategy and a domain adversarial network, and using the target spatial transcriptome data to obtain a trained feature extraction model for extracting embedding information and a prediction model for predicting cell type proportions includes: Inputting the target spatial transcriptome data into the pre-constructed encoder for training the extraction of embedding information; Optimizing the spatial information in the embedding information based on a contrast learning strategy to determine the mutual relationship between spatial sites, and calculating the corresponding contrast loss; Inputting the embedding information into the domain adversarial network for domain adversarial training to calculate the corresponding binary cross-entropy loss, and inputting the embedding information into the initial prediction model for prediction training of cell type proportions, and calculating the corresponding L1 loss; Until the contrast loss, the binary cross-entropy loss, and the L1 loss satisfy the preset loss condition, a trained feature extraction model for extracting the embedding information and a prediction model for predicting the cell type ratio are obtained.

7. The method for constructing a deconvolution model according to claim 6, wherein During the process of predicting the cell type ratio through the prediction model, it further includes: Processing the high-dimensional embedding representation generated by the feature extraction model according to preset neural network operations to generate a probability distribution of the corresponding cell type, and predicting the cell type ratio based on the probability distribution of the cell type; wherein, the preset neural network operations include: linear transformation operation, Leaky ReLU activation function, Dropout operation, and softmax function.

8. An anti-convolution model construction device, characterized in that, The device includes: A data simulation module for generating simulated spatial transcriptome data with cell type ratio labels based on single-cell transcriptome data of the same tissue. A data sampling module for obtaining real spatial transcriptome data sampled based on a 3D concentric sphere sampling strategy. A gene screening module for screening differentially expressed genes based on the simulated spatial transcriptome data and the real spatial transcriptome data to obtain target spatial transcriptome data containing differentially expressed genes. A model training module for training a pre-constructed encoder and an initial prediction model based on a contrast learning strategy and a domain adversarial network, and using the target spatial transcriptome data to obtain a trained feature extraction model for extracting embedding information and a prediction model for predicting the cell type ratio; wherein, during the training process, the embedding information output by the encoder is used as the input of the initial prediction model and the domain adversarial network.

9. A terminal, characterized in that, It includes: A memory, a processor, and a deconvolution model construction program stored on the memory and executable on the processor. When the deconvolution model construction program is executed by the processor, it implements the steps of the deconvolution model construction method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that can be executed to implement the steps of the deconvolution model construction method according to any one of claims 1 to 7.

Images