A cell proportion prediction method and device fusing prior biological knowledge

Abstract

The application discloses a cell proportion prediction method and device fusing prior biological knowledge, and the method comprises the following steps: inputting mixed RNA expression data of a to-be-tested tissue sample into a pre-trained double-flow deconvolution model, wherein the double-flow deconvolution model comprises a feature extraction module and a feature fusion prediction module, the feature extraction module comprises two parallel branches: a deep neural network and a knowledge-driven sparse neural network, and the data processing process comprises the following steps: inputting the mixed RNA expression data into the deep neural network to extract a data-driven hidden layer high-dimensional feature vector; inputting the mixed RNA expression data into the knowledge-driven sparse neural network to extract a knowledge feature vector conforming to biological rules by using a pre-constructed knowledge mask; and inputting the hidden layer high-dimensional feature vector and the knowledge feature vector into a full connection layer and a Softmax activation function after splicing in the feature fusion prediction module, so as to obtain a prediction result of the proportion of each type of cell in the to-be-tested tissue sample.

Description

Technical Field

[0001] This application relates to a method and apparatus for predicting cell proportions by integrating prior biological knowledge, belonging to the fields of bioinformatics and artificial intelligence. Background Technology

[0002] Understanding the precise cellular composition within complex tissues is fundamental to modern biomedicine and crucial for deciphering mechanisms of human development, immune regulation, and disease progression. While traditional bulk RNA sequencing (bulk RNA-seq) enables gene expression profiling of large cohorts, it fundamentally measures the average transcriptional output of a mixed cell population. This "averaging effect" masks cell-specific signals, making it difficult to distinguish between changes in gene regulation and changes in cell abundance. Although single-cell RNA sequencing (scRNA-seq) can capture the transcriptomic landscape at cellular resolution, its high cost and technical noise limitations restrict its application in large-scale epidemiological studies and routine clinical diagnosis.

[0003] Computational cell deconvolution has emerged as a cost-effective alternative. Current deconvolution methods can be broadly categorized into reference-free and reference-based methods. Reference-based methods, such as CIBERSORTx and BayesPrism, typically assume a linear relationship between gene expression and cell proportion. However, gene expression in biological systems is inherently random and nonlinear, influenced by confounding factors such as cell interactions and metabolic states. Summary of the Invention

[0004] Objective: In view of at least one of the above technical problems, this application provides a method and apparatus for predicting cell proportions by incorporating prior biological knowledge.

[0005] The technical solution adopted in this application is:

[0006] Firstly, this application provides a cell proportion prediction method that integrates prior biological knowledge, including:

[0007] Obtain mixed RNA expression data from the tissue sample to be tested;

[0008] The mixed RNA expression data is input into a pre-trained two-stream deconvolution model to obtain the predicted proportion of each type of cell in the tissue sample to be tested.

[0009] The dual-stream deconvolution model includes a feature extraction module and a feature fusion prediction module. The feature extraction module includes two parallel branches: a deep neural network and a knowledge-driven sparse neural network. The data processing includes:

[0010] The mixed RNA expression data is input into a deep neural network to extract data-driven high-dimensional feature vectors of the hidden layers;

[0011] The mixed RNA expression data is input into a knowledge-driven sparse neural network. The network is then structurally constrained using a pre-constructed knowledge mask to extract knowledge feature vectors that conform to biological laws.

[0012] In the feature fusion prediction module, the hidden layer high-dimensional feature vector and the knowledge feature vector are concatenated and then input into a fully connected layer and a Softmax activation function to obtain the prediction results of the proportion of each type of cell in the tissue sample to be tested.

[0013] In some embodiments, the method for constructing the knowledge mask includes:

[0014] To obtain information on known biological pathways and protein-protein interaction networks;

[0015] The biological pathway information is encoded into a binary gene-pathway matrix;

[0016] Based on the protein-protein interaction network information and the corresponding confidence level, a protein-protein interaction adjacency matrix with confidence level as the weight is constructed by filtering using a confidence level threshold.

[0017] The knowledge mask is generated by concatenating the gene-pathway matrix and the protein-protein interaction adjacency matrix along the feature dimension.

[0018] Secondly, this application provides a cell ratio prediction device that integrates prior biological knowledge, including a processor and a storage medium;

[0019] The storage medium is used to store instructions;

[0020] The processor is configured to operate according to the instructions to execute the method according to the first aspect.

[0021] Thirdly, this application provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in the first aspect.

[0022] Fourthly, this application provides a computer device including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the method described in the first aspect.

[0023] Beneficial Effects: The cell proportion prediction method and device integrating prior biological knowledge provided in this application have the following advantages: Breaking the "black box" limitation and enhancing biological consistency: This application innovatively proposes a knowledge-driven sparse neural network, which constrains the network topology through knowledge masks (biological pathways and PPIs). This forces the model to rely not only on statistical correlations but also on known biological interaction logics when extracting features, thereby effectively filtering out noise from specific datasets and capturing truly biologically meaningful features.

[0024] Significantly improved prediction accuracy and robustness: A dual-stream architecture is adopted to integrate data-driven hidden features and knowledge-driven structural features. In cross-validation across multiple independent cohorts (such as peripheral blood and ovarian cancer tumor microenvironment), this application demonstrates extremely high accuracy in deconvolution of highly similar cell subpopulations (such as CD4 and CD8 T cells), overcoming batch effects.

[0025] Multi-layered model interpretability: This application incorporates SHAP-based interpretability units, enabling dual interpretation at the gene level and knowledge level (pathway or protein-protein interaction PPI nodes). Researchers can clearly quantify which marker genes or biological pathways drive the prediction of specific cell proportions, bridging computational inference with clinical a priori intuition, greatly enhancing the model's credibility and practical value in clinical and mechanistic studies. Attached Figure Description

[0026] Figure 1 This is a flowchart illustrating a cell proportion prediction method that integrates prior biological knowledge according to an embodiment of this application.

[0027] Figure 2 This is a schematic diagram of the network structure according to the embodiments of this application;

[0028] Figure 3 This is a schematic diagram of the process for constructing pseudo-mixed sample training data in an embodiment of this application;

[0029] Figure 4 This is a schematic diagram of the interpretability unit logic in the embodiment of this application, which incorporates the SHAP (Shapley Additive exPlanations) framework.

[0030] Figure 5 This is a schematic diagram comparing the prediction results of the consistency correlation coefficient (CCC) of the method in this application with those of other existing methods.

[0031] Figure 6 This is a schematic diagram comparing the mean absolute error (MAE) prediction results of the method in this application with those of other existing methods.

[0032] Figure 7This is a schematic diagram illustrating the interpretability results of the method in the embodiments of this application. Detailed Implementation

[0033] The present application will be further described below with reference to the accompanying drawings and embodiments. The following embodiments are only used to more clearly illustrate the technical solutions of the present application, and should not be used to limit the scope of protection of the present application.

[0034] In the description of this application, "several" means one or more, "multiple" means two or more, "greater than," "less than," and "exceeding" are understood to exclude the stated number, while "above," "below," and "within" are understood to include the stated number. The use of "first" and "second" in the description is merely for distinguishing technical features and should not be construed as indicating or implying relative importance, or implicitly indicating the number of indicated technical features, or implicitly indicating the order of the indicated technical features.

[0035] In the description of this application, the terms "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples," etc., refer to specific features, structures, materials, or characteristics described in connection with that embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0036] The term "and / or" simply describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the character " / " generally indicates that the preceding and following related objects have an "or" relationship.

[0037] Deep learning-based methods (such as Scaden and TAPE) utilize deep neural networks to approximate the complex nonlinear mapping between gene expression profiles and cellular components, significantly improving robustness against technical noise in transcriptome data. Despite the improved prediction accuracy achieved by deep learning, mainstream deconvolutional frameworks face a critical bottleneck: a lack of interpretability (black-box models). Standard deep neural networks treat the transcriptome as an unstructured vector space, ignoring the hierarchical structure of biological systems (i.e., gene expression is meticulously orchestrated by tightly regulated signal cascades and interaction networks). Lacking these inductive biases, purely data-driven models are prone to overfitting to technical artifacts, learning the noise of the sequencing platform rather than cellular identity signals. Even when some models attempt to incorporate pathway information, they often employ simple element-wise multiplication, failing to capture the inherent collaborative interactions and directional logic within biological networks. The lack of transparent decision-making processes prevents clinicians and biologists from verifying whether model predictions are based on known biological mechanisms, severely limiting their practicality in mechanistic research. To address this, this application provides a cell proportion prediction method that integrates prior biological knowledge, which can overcome the shortcomings of existing deep learning deconvolution methods that rely solely on pure data-driven approaches, resulting in a "black box" problem, easy overfitting to sequencing noise, and a lack of biological interpretability in the prediction results.

[0038] Example 1: This example provides a cell proportion prediction method that integrates prior biological knowledge, such as... Figure 1 As shown, it includes:

[0039] Obtain mixed RNA expression data from the tissue sample to be tested;

[0040] The mixed RNA expression data is input into a pre-trained two-stream deconvolution model to obtain the predicted proportion of each type of cell in the tissue sample to be tested.

[0041] The dual-stream deconvolution model includes a feature extraction module and a feature fusion prediction module. The feature extraction module includes two parallel branches: a deep neural network (DNN) and a knowledge-based sparse neural network (KSNN). The data processing includes:

[0042] The mixed RNA expression data is input into a deep neural network to extract data-driven high-dimensional feature vectors of the hidden layers;

[0043] The mixed RNA expression data is input into a knowledge-driven sparse neural network. The network is then structurally constrained using a pre-constructed knowledge mask to extract knowledge feature vectors that conform to biological laws.

[0044] In the feature fusion prediction module, the hidden layer high-dimensional feature vector and the knowledge feature vector are concatenated and then input into a fully connected layer and a Softmax activation function to obtain the prediction results of the proportion of each type of cell in the tissue sample to be tested.

[0045] In some embodiments, the method for constructing the knowledge mask includes:

[0046] Obtain information on known biological pathways and protein-protein interaction (PPI) networks;

[0047] The biological pathway information is encoded into a binary gene-pathway matrix;

[0048] Based on the protein-protein interaction network information and the corresponding confidence level, a protein-protein interaction adjacency matrix with confidence level as the weight is constructed by filtering using a confidence level threshold.

[0049] The knowledge mask is generated by concatenating the gene-pathway matrix and the protein-protein interaction adjacency matrix along the feature dimension.

[0050] In this embodiment, biological pathway information is obtained from MSigDB databases (such as KEGG, Reactome, etc.), filtered, and encoded into a binary gene-pathway matrix. Simultaneously, protein-protein interaction network information is obtained from the STRING database, and protein-protein interactions with a confidence level not less than a confidence threshold are retained to construct a protein-protein interaction adjacency matrix weighted by confidence level.

[0051] The gene-pathway matrix and the protein-protein adjacency matrix are concatenated along the feature dimension to form the final binary / weighted knowledge mask, which is represented as follows: , For the number of genes, This represents the number of knowledge features.

[0052] In some embodiments, such as Figure 2 As shown, the deep neural network (DNN) consists of five fully connected layers with decreasing dimensions (512, 256, 128, 64, and 32 neurons respectively), and uses conventional deep learning to extract mixed RNA expression data end-to-end. Hidden layer high-dimensional feature vectors , This represents a deep neural network.

[0053] In some embodiments, such as Figure 2 As shown, the knowledge-driven sparse neural network (KSNN) includes a prior knowledge sparse layer and multiple fully connected layers connected in sequence;

[0054] In this context, the connections between neurons in the prior knowledge sparse layer are restricted by the knowledge mask, allowing connections only between nodes with known biological interaction relationships.

[0055] It should be noted that in the aforementioned prior knowledge sparse layer, node connections are not fully connected, but rather determined by a constructed knowledge mask. The definition is to extract only knowledge features with biological basis and filter out spurious relevance.

[0056] The knowledge-driven sparse neural network processing process is represented as follows:

[0057] ;

[0058] in, For knowledge feature vectors, For the input mixed RNA expression data, This represents a knowledge-driven sparse neural network. For learnable weight matrix, As a knowledge mask, This represents the Hadama product.

[0059] In a specific embodiment, the present invention constructs the knowledge mask by integrating multidimensional biological prior knowledge. First, 1,603 high-quality biological pathways were selected from the MSigDB database (covering KEGG, Reactome, PID, and BioCarta, etc.) and encoded into a binary gene-pathway matrix. Second, protein-protein interaction network information was obtained from the STRING database, and only reliable connections with a confidence score of not less than 0.7 were retained to construct a protein-protein interaction adjacency matrix. A knowledge mask was then constructed by concatenating the gene-pathway matrix and the protein-protein interaction adjacency matrix along their feature dimensions. This is used to impose structured constraints on the first layer (i.e., the prior knowledge sparse layer) of the knowledge-driven sparse neural network.

[0060] The input neurons of the prior knowledge sparse layer correspond one-to-one with the input gene dimension, and the output neurons correspond one-to-one with the knowledge feature dimension, so that each output neuron represents a specific biological functional module (such as a biological pathway or a protein-protein interaction node). The learnable weight matrix is ​​forced through the Hadamard product. The connection weights, which lack biological basis, are reset to 0, causing the nodes in this layer to not use full connections, but rather be strictly subject to knowledge masks. The defined structural constraints ensure that the information flow of the two-stream deconvolution model is only constrained by known biological laws. After this prior knowledge sparse layer, five fully connected layers with decreasing dimensions (512, 256, 128, 64, and 32 neurons respectively) are further configured for high-order feature extraction.

[0061] In this embodiment, in the feature fusion prediction module, the hidden layer high-dimensional feature vector is... With the knowledge feature vector After concatenation, the image is mapped to the output space through a fully connected layer. The Softmax activation function is then used to obtain the predicted proportions of each cell type in the test tissue sample, as shown below:

[0062] ;

[0063] in, For the first The true value of the proportion of cell-like organisms. This represents the Softmax activation function. Indicates splicing, , These are the weight matrix and bias of the fully connected layer, respectively.

[0064] In some embodiments, the training method of the two-stream deconvolution model includes:

[0065] S100. Construct the training dataset;

[0066] Supervised training requires a large number of samples with known cell proportions, but real mixed RNA sequencing data often lacks such real labels. This embodiment employs a strategy of simulating pseudo-mixed samples based on single-cell data.

[0067] In this embodiment, as Figure 3 As shown, S100 specifically includes:

[0068] S101. Obtain single-cell RNA sequencing reference data and cell type labeling;

[0069] Generate target cell scale vector for each simulated sample ,satisfy , This is the target cell type proportion vector. This represents the total number of cell types. Sampling can be performed using a Dirichlet distribution.

[0070] S102. The number of cells sampled is determined by simulating technical variations in the sequencing process using a Poisson distribution, expressed as:

[0071] ;

[0072] in, For the number of cells sampled, Indicates cell type Expected cell count The Poisson distribution, , To preset the total library size, This is the target cell type ratio vector.

[0073] S103. Based on the determined number of cell samples, sample with replacement from the single-cell RNA sequencing reference data, accumulate the expression profiles of the sampled cells, add Gaussian noise, and generate simulated pseudo-mixed RNA samples and corresponding real cell proportion labels, which serve as the training dataset for the two-stream deconvolution model.

[0074] S200. The two-stream deconvolution model is trained using the training dataset to obtain a trained two-stream deconvolution model.

[0075] The two-stream deconvolution model employs a loss function combining MSE and MAE during training; in some embodiments, the loss function of the two-stream deconvolution model is expressed as:

[0076] ;

[0077] in, For loss function, For the first Predicted cell proportions For the first The true value of the proportion of cell-like organisms. This represents the total number of cell types.

[0078] The Adam optimization algorithm is used to update the network parameters of the two-stream deconvolution model. To improve prediction robustness, multiple independent models can be trained in parallel, and the final prediction ratio is determined by the average of the outputs of each model.

[0079] Furthermore, to overcome the "black box" limitation and integrate the SHAP framework into the inference pipeline, this application introduces SHAP to achieve model interpretability, including:

[0080] The trained two-stream deconvolutional model is input into the SHAP interpretable analysis unit, and feature attribution is calculated based on the SHAP framework, such as... Figure 4 As shown:

[0081] At the gene level, the SHAP value of the input gene relative to the predicted proportion of each cell type (the feature after concatenating the hidden layer high-dimensional feature vector and the knowledge feature vector) is calculated to identify the marker gene that drives the prediction of a specific cell type.

[0082] It should be noted that this step, which determines whether high expression of a specific gene drives the prediction of a particular cell type, either positively or negatively. For example... Figure 7 As shown, the SHAP value using biological knowledge is an order of magnitude larger than that without biological knowledge. This indicates that knowledge integration transforms classic markers (e.g., CD19, CD79A, PAX5) into powerful, decisive predictors, whereas without biological knowledge, the model can only rely on weak, spurious signals (e.g., CYP1B1, which is completely unrelated to B cells). The magnitude of influence of biological modules (pathways or PPI sets) is ranked.

[0083] At the knowledge level, the average absolute SHAP value of the knowledge feature vector is calculated to quantify the contribution of each biological pathway and protein-protein interaction node to the overall decision logic of the model.

[0084] Simulation Experiment: Construction and Validation of Peripheral Blood Mononuclear Cell Dataset

[0085] The peripheral blood mononuclear cell dataset was obtained from the PBMC8k dataset in the 10x Genomics official database. Following the strategy for constructing the training dataset in Example 1, 8,000 pseudo-mixed samples with known proportion labels were generated, containing 6 major cell types (B cells, CD4 T cells, CD8 T cells, mononuclear cells, natural killer cells, and unknown cells), which served as the training dataset for this example.

[0086] To verify the model's capabilities in real-world mixed RNA expression data environments, the trained model was applied to eight independent peripheral blood mononuclear mixed RNA expression datasets, including Sdy67, Sdy67-250, GSE107011, GSE65133, GSE107572, GSE120502, GSE107990, and GSE193141. Sdy67 and Sdy67-250 were obtained from the ImmPort database, while the remaining data were obtained from the Gene Expression Omnibus (GEO) database.

[0087] The flow cytometry measurements associated with each dataset were used as the true values ​​for comparison. The method provided in this application outperforms existing baseline models such as DeSide, TAPE, Scaden (including TAPE-O and TAPE-H), ReDeconv, BayesPrism, and CIBERSORTx in terms of prediction accuracy consistency correlation coefficient (CCC) and mean absolute error (MAE). Figure 5 and Figure 6The figures show the consistent correlation coefficients (higher is better) and mean absolute errors (MAE, lower is better) for six cell types in eight real peripheral blood mononuclear mixed RNA expression datasets based on embodiments of this application. Notably, embodiments of this application lead in lymphocyte quantification (B cells, CD4 T cells, CD8 T cells) and also outperform most baseline models in the quantification of the other three cell types.

[0088] Example 2: Based on Example 1, this example provides a cell ratio prediction device that integrates prior biological knowledge, including a processor and a storage medium;

[0089] The storage medium is used to store instructions;

[0090] The processor is configured to operate according to the instructions to execute the method according to Embodiment 1.

[0091] Example 3: Based on Example 1, this example provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described in Example 1.

[0092] Example 4: Based on Example 1, this example provides a computer device, including a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the method described in Example 1.

[0093] Example 5: Based on Example 1, this example provides a computer program product, including a computer program that, when executed by a processor, implements the method described in Example 1.

[0094] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0095] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, as well as combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0096] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0097] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0098] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the principle of this application, and these improvements and modifications should also be considered within the scope of protection of this application.

Claims

1. A method for predicting cell proportions by integrating prior biological knowledge, characterized in that, include: Obtain mixed RNA expression data from the tissue sample to be tested; The mixed RNA expression data is input into a pre-trained two-stream deconvolution model to obtain the predicted proportion of each type of cell in the tissue sample to be tested. The dual-stream deconvolution model includes a feature extraction module and a feature fusion prediction module. The feature extraction module includes two parallel branches: a deep neural network and a knowledge-driven sparse neural network. The data processing includes: The mixed RNA expression data is input into a deep neural network to extract data-driven high-dimensional feature vectors of the hidden layers; The mixed RNA expression data is input into a knowledge-driven sparse neural network. The network is then structurally constrained using a pre-constructed knowledge mask to extract knowledge feature vectors that conform to biological laws. In the feature fusion prediction module, the hidden layer high-dimensional feature vector and the knowledge feature vector are concatenated and then input into the fully connected layer and the Softmax activation function to obtain the prediction results of the proportion of each type of cell in the tissue sample to be tested. The method for constructing the knowledge mask includes: To obtain information on known biological pathways and protein-protein interaction networks; The biological pathway information is encoded into a binary gene-pathway matrix; Based on the protein-protein interaction network information and the corresponding confidence level, a protein-protein interaction adjacency matrix with confidence level as the weight is constructed by filtering using a confidence level threshold. The knowledge mask is generated by concatenating the gene-pathway matrix and the protein-protein interaction adjacency matrix along the feature dimension. The knowledge-driven sparse neural network includes a prior knowledge sparse layer and multiple fully connected layers connected in sequence. In this context, the connections between neurons in the prior knowledge sparse layer are restricted by the knowledge mask, allowing connections only between nodes with known biological interaction relationships. The knowledge-driven sparse neural network processing process is represented as follows: ； in, For knowledge feature vectors, For the input mixed RNA expression data, For learnable weight matrix, As a knowledge mask, This represents the Hadama product.

2. The method according to claim 1, characterized in that, The training method for the two-stream deconvolution model includes: The training dataset was constructed by sampling from single-cell RNA sequencing reference data; The two-stream deconvolution model is trained using the training dataset to obtain a trained two-stream deconvolution model.

3. The method according to claim 2, characterized in that, The training dataset was constructed by sampling from single-cell RNA sequencing reference data, including: Obtain single-cell RNA sequencing reference data and cell type labeling; The number of cells to be sampled was determined by simulating technical variations during the sequencing process using a Poisson distribution. Based on the determined number of cells sampled, sampling with replacement is performed from the single-cell RNA sequencing reference data. The expression profiles of the sampled cells are accumulated, and Gaussian noise is added to generate simulated pseudo-mixed RNA samples and corresponding real cell proportion labels, which serve as the training dataset for the two-stream deconvolution model.

4. The method according to claim 3, characterized in that, The number of cells sampled is determined by simulating technical variations in the sequencing process using a Poisson distribution, denoted as: ； in, For the number of cells sampled, Indicates cell type Expected cell count The Poisson distribution, , To preset the total library size, This is the target cell type ratio vector.

5. The method according to claim 1, characterized in that, The loss function of the two-stream deconvolution model is expressed as: ； in, For loss function, For the first Predicted cell proportions For the first The true value of the proportion of cell-like organisms. This represents the total number of cell types.

6. A cell ratio prediction device integrating prior biological knowledge, characterized in that, Including processor and storage media; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the method according to any one of claims 1 to 5.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method according to any one of claims 1 to 5.

8. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the method according to any one of claims 1 to 5.

Images