Joint deconvolution domain identification and alignment method and system, terminal and medium

Abstract

The invention relates to the technical field of data processing, and discloses a combined deconvolution domain recognition and alignment method and system, a terminal and a medium, and the combined deconvolution domain recognition and alignment method comprises the steps: obtaining multi-slice gene transcriptome data and reference data; variational distribution is obtained according to the multi-slice gene transcriptome data; performing joint optimization on the variation distribution according to the multi-slice gene transcriptome data and the reference data to obtain a target parameter; and generating a deconvolution result, a spatial domain recognition result and unified representation information according to the target parameters. According to the method, semantic drift caused by a step-by-step process can be eliminated, so that a deconvolution result, domain division and alignment are collaboratively optimized in the same representation, and the efficiency and accuracy of data processing are improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of data processing, in particular to a method and system for joint deconvolution, domain identification and alignment, a terminal and a medium. BACKGROUND

[0002] Spatial resolved transcriptomics (SRT) is a sequencing technology that simultaneously captures gene expression profiles and spatial coordinates on a single tissue section, providing a new means for analyzing the molecular characteristics, subpopulation distribution, and interactions of cells in situ. Unlike traditional single-cell sequencing, SRT preserves the key dimension of "spatial location of expression" while obtaining transcription information, showing high application value in tumor microenvironment analysis, developmental trajectory tracking, and drug target discovery. According to the experimental principle, it can be roughly divided into two systems: optical imaging-in situ hybridization and barcode capture-high-throughput sequencing. The barcode route first lays a spatial coding array on the section surface, captures mRNA in situ, and then transfers the whole to the laboratory for subsequent amplification and next-generation sequencing. This system can detect 20-30 thousand genes in one experiment and cover several square centimeters of tissue area, combining high throughput and lower cost, and is the most widely used solution in clinical and industrial applications. However, due to the diameter and diffusion radius of the barcode, most capture points still correspond to "multi-cell pixels". For example, on the Visium (55 um) or Slide-seqV2 (10 um) matrix, a pixel often contains 2-20 cells. Pixel mixing makes it difficult to directly distinguish the boundaries between immune infiltrating cells and tumor cells, and also hinders the analysis of the interaction between stroma, immune, and tumor. It is necessary to use deconvolution methods to decompose and restore the mixed gene expression signals of each measurement point to several single-cell or cell-type expression profiles and composition ratios to provide interpretable information for clinical diagnosis and treatment.

[0003] However, for spatial transcriptome data analysis, cell type deconvolution, tissue spatial domain identification, and cross-section integration are usually implemented separately and decoupled using different tools, that is, in the existing process, deconvolution, clustering, and alignment are run independently, lacking joint optimization in the same interpretable latent space, resulting in semantic drift and error accumulation.

[0004] Therefore, the prior art still needs to be improved and developed. SUMMARY

[0005] The main purpose of the present application is to provide a method and system for joint deconvolution, domain identification, and alignment, a terminal and a medium, aiming to solve the problem of semantic drift and error accumulation caused by independent implementation of deconvolution, spatial domain clustering, and multi-section integration in the process of analyzing spatial transcriptome data in the prior art.

[0006] The first aspect of the embodiment of the present application provides a method for joint deconvolution domain identification and alignment, which comprises the following steps: obtaining multi-slice gene transcriptome data and reference data; obtaining a variational distribution according to the multi-slice gene transcriptome data; jointly optimizing the variational distribution according to the multi-slice gene transcriptome data and the reference data to obtain a target parameter; generating a deconvolution result, a spatial domain identification result and unified representation information according to the target parameter.

[0007] Optionally, in an embodiment of the present application, the multi-slice gene transcriptome data comprises gene expression count matrix data of multiple slices; The obtaining a variational distribution according to the multi-slice gene transcriptome data specifically comprises: constructing a hierarchical topic model according to the gene expression count matrix data; inference is performed on the hierarchical topic model to obtain a variational distribution.

[0008] Optionally, in an embodiment of the present application, the multi-slice gene transcriptome data further comprises spatial coordinate data of multiple slices; The jointly optimizing the variational distribution according to the multi-slice gene transcriptome data and the reference data to obtain a target parameter specifically comprises: obtaining a lower bound of evidence according to the variational distribution; optimizing the lower bound of evidence according to the spatial coordinate data and the reference data to obtain a target parameter.

[0009] Optionally, in an embodiment of the present application, the reference data is single-cell transcriptome data; The optimizing the lower bound of evidence according to the spatial coordinate data and the single-cell transcriptome data to obtain a target parameter specifically comprises: optimizing the lower bound of evidence to obtain estimated data; obtaining a component representation according to the estimated data; constructing a neighborhood graph according to the spatial coordinate data, and obtaining a target parameter according to the neighborhood graph, the single-cell transcriptome data and the component representation.

[0010] Optionally, in an embodiment of the present application, the obtaining a target parameter according to the neighborhood graph, the single-cell transcriptome data and the component representation specifically comprises: obtaining a global optimization target according to the neighborhood graph, the single-cell transcriptome data, and the composition representation; determining a target parameter according to the global optimization target.

[0011] Optionally, in an embodiment of the present application, the obtaining a composition representation according to the estimated data specifically comprises: obtaining a cell type composition vector according to the estimated data; obtaining a to-be-aligned representation according to the cell type composition vector, and performing cross-slice alignment on the to-be-aligned representation to obtain the composition representation.

[0012] Optionally, in an embodiment of the present application, the jointly optimizing the variational distribution according to the multi-slice gene transcriptome data and the reference data to obtain the target parameter further comprises: constructing an initial model, obtaining training data, and training the initial model according to the training data to obtain a trained initial model; inputting the multi-slice gene transcriptome data into the trained initial model to generate an initial parameter of the variational distribution.

[0013] The second aspect of the embodiments of the present application further provides a system for jointly deconvolving domain identification and alignment, wherein the system for jointly deconvolving domain identification and alignment is applied to the method for jointly deconvolving domain identification and alignment in any of the above solutions; and the system for jointly deconvolving domain identification and alignment comprises: a data acquisition module configured to acquire multi-slice gene transcriptome data and reference data; a variational inference module configured to obtain a variational distribution according to the multi-slice gene transcriptome data; a joint optimization module configured to jointly optimize the variational distribution according to the multi-slice gene transcriptome data and the reference data to obtain a target parameter; a result output module configured to generate a deconvolution result, a spatial domain identification result, and unified representation information according to the target parameter.

[0014] The third aspect of the embodiments of the present application further provides a terminal, wherein the terminal comprises a memory, a processor, and a program for jointly deconvolving domain identification and alignment stored in the memory and executable on the processor, and the program for jointly deconvolving domain identification and alignment is executed by the processor to implement the steps of the method for jointly deconvolving domain identification and alignment.

[0015] The fourth aspect of the embodiment of the application further provides a computer readable storage medium, wherein the computer readable storage medium stores a joint deconvolution domain identification and alignment program, and the joint deconvolution domain identification and alignment program is executed by a processor to implement the steps of the joint deconvolution domain identification and alignment method.

[0016] Beneficial effects: the application provides a joint deconvolution domain identification and alignment method, system, terminal and medium, the application models gene expression data by a topic model, obtains a low-dimensional interpretable core representation composed of cell types (topics), directly performs spatial domain division and cross-slice distribution registration on the unified embedding, realizes spatial consistency through graph constraints, and adopts a linearly expandable calculation strategy to adapt to large-scale multi-slice data, thereby eliminating semantic drift caused by a step-by-step process, making the deconvolution result, domain division and alignment be optimized in the same representation, and improving the efficiency and accuracy of data processing. BRIEF DESCRIPTION OF DRAWINGS

[0017] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiment or prior art description. Obviously, the drawings in the following description are only some embodiments of the application, and for those skilled in the art, other drawings can also be obtained without creative labor.

[0018] Figure 1 a flowchart of a preferred embodiment of the joint deconvolution domain identification and alignment method of the application; Figure 2 a schematic diagram of integration, clustering and deconvolution model of spatial transcriptome on multiple slices in a preferred embodiment of the joint deconvolution domain identification and alignment method of the application; Figure 3 a schematic diagram of performance effect on 10x Visium human brain dorsal prefrontal cortex data in a preferred embodiment of the joint deconvolution domain identification and alignment method of the application; Figure 4 a schematic diagram of performance effect on Stereo-seq mouse embryo data in a preferred embodiment of the joint deconvolution domain identification and alignment method of the application Figure 5 a structure diagram of a preferred embodiment of the joint deconvolution domain identification and alignment system of the application; Figure 6 a structure diagram of a preferred embodiment of the terminal of the application.

[0019] Explanation of reference signs: 100, data acquisition module; 200, variational inference module; 300, joint optimization module; 400, result output module. Detailed Implementation

[0020] To make the objectives, technical solutions, and effects of this application clearer and more explicit, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. The described embodiments are only possible technical implementations of this application and not all possible implementations. Based on the embodiments in this application, those skilled in the art can obtain other embodiments without creative effort, and these embodiments are also within the protection scope of this application.

[0021] First, let's introduce the terms used in the embodiments of this application: SRT, Spatial Resolved Transcriptomics, acquires information on gene expression and spatial coordinates simultaneously from tissue slices, which is used to analyze the expression patterns and structures of cells / tissues in the spatial dimension. scRNA-seq, Single Cell RNA Sequencing, is a technique that sequences the transcriptome of a single cell to obtain a cell-level gene expression profile, often used as a reference for spatial deconvolution or cell type annotation. OT, Optimal Transport, is a matching / mapping technique that finds the minimum "transportation cost" between two distributions. It is often used for distribution alignment, batch correction, and trajectory inference across slices / platforms. UMI, Unique Molecular Identifier, is a short barcode added to each molecule during the library construction stage. It is used to remove PCR amplification duplications and improve counting accuracy. It is commonly used in spatial / single-cell sequencing. GNN, Graph Neural Network, is a deep model that performs message passing and representation learning on a graph structure (nodes / edges). It often uses spatially adjacent points as a graph for aggregation, smoothing, or clustering. WGAN, or Wasserstein Generative Adversarial Network, is a GAN architecture that uses Wasserstein-1 distance as the adversarial target. It uses a "critic" as the discriminator, which has better stability and metric significance, and is often used for distribution registration and alignment. VAE, Variational Autoencoder, is a generative model that combines deep networks with variational inference. It learns latent representations by maximizing ELBO and is often used for dimensionality reduction, denoising, and generative modeling. ELBO, Evidence Lower Bound, the objective function of Variational Inference, is a lower bound of log marginal likelihood, used for training VAE, topic model, etc. MRF, Markov Random Field, a probabilistic graphical model on undirected graph, characterizes the consistency of neighborhood with potential function, often used for spatial smoothing / boundary refinement. BIC, Bayesian Information Criterion, a model selection criterion, often used for selecting hyperparameters like the number of clusters / number of topics. PCA, Principal Component Analysis, a classical linear dimensionality reduction method, projects high-dimensional data onto a set of orthogonal principal components with maximum variance, used for denoising, visualization, and as input for downstream analysis. UMAP, Uniform Manifold Approximation and Projection, a nonlinear dimensionality reduction / visualization method, based on manifold learning and fuzzy topology, preserves local neighborhood structure while trying to maintain global shape, often used for 2D / 3D visualization of high-dimensional expression data. ARI, Adjusted Rand Index, a clustering consistency evaluation index, measures the matching degree of two partitions (such as predicted clusters and true labels), and corrects for random consistency, with a value range of [-1, 1], with a larger value indicating greater consistency. iLIST, integration Local Inverse Simpson’s Index, a batch mixing degree / integration quality index: in low-dimensional embedding or kNN graph, calculate the effective number of "batch origin" within the local neighborhood (reflecting the mixing degree); the higher the value, the more fully mixed the different batches in the neighborhood, used to evaluate the cross-batch / cross-slice integration effect.

[0022] In related technologies, for the analysis of spatial transcriptome data, cell type deconvolution, tissue spatial domain recognition, and cross-slice integration are usually implemented separately and decoupled using different tools. These four points are as follows: (1) In terms of deconvolution, the mainstream approach is to decompose each spatial site into the proportion of multiple cell types using single-cell RNA sequencing (scRNA-seq) or marker genes as a reference. The common problem of such methods is: strong dependence on reference and batch consistency, systematic bias when reference and target sample do not match; usually single slice is independently estimated, cannot utilize the shared structure across slices while estimating cell proportion; they are often independent of subsequent spatial domain recognition / integration steps, and cannot be optimized in a consistent latent space.

[0023] (2) In terms of spatial domain recognition, existing methods often cluster on high-variation genes or low-dimensional representations of gene expression. This approach has common limitations: clustering results are not equivalent to cell type composition differences, and lack consistency with deconvolution; many methods require pre-setting the number of clusters K and are sensitive to hyperparameters; when extended to multiple slices / large-scale data, the computational and memory overhead of constructing the field graph grows rapidly with the number of sites and graph edges.

[0024] (3) In terms of cross-slice integration / alignment, existing work such as optimal transport aligns expression and physical distance to align adjacent slices, reconstructs 3D tissue structure through deep representation, projects different slices into a common space or reconstructs quasi-three-dimensional structure. The limitations of such methods are: ignoring the more stable and interpretable representation of "cell type composition"; when the batch or tissue heterogeneity of the slices is strong, the alignment between expression distributions is easily overcorrected; the computational complexity of full graph matching or optimal transport (OT) across slices is high.

[0025] (4) In terms of dimensionality reduction / embedding representation, existing techniques often construct low-dimensional representations from gene expression for visualization and subsequent tasks, but such representations are usually independent of the deconvolution results: deconvolution gives cell proportions, while clustering / alignment relies on another set of expression embeddings, which are not constrained by each other, making it difficult to ensure consistency in "domain boundary-cell composition change-cross slice alignment".

[0026] In summary, the main shortcomings of existing technologies are: 1) Deconvolution, clustering, and alignment in existing processes are run independently, lacking joint optimization in the same interpretable latent space, causing semantic drift and error accumulation. 2) The construction and training of cross-slice spatial field graphs result in excessive memory and time overhead, making it difficult to linearly extend to multiple slices and large-scale data. 3) Clustering based on expression and real cell type proportion may have semantic inconsistency, and domain boundaries are often not aligned with cell type distribution. 4) A large amount of manual setting (such as cluster number, hyperparameter) is required, with repeated calculations and insufficient stability when migrating to heterogeneous platforms. 5) Cross-slice integration is often based on gene expression, which is easily affected by batch and platform differences and is not robust to cell composition.

[0027] The present application aims at the problems in the prior art that deconvolution, spatial domain clustering and multi-slice integration are mutually disjointed, cross-slice integration is not robust, and large-scale data calculation expansion is limited. The present application simultaneously completes the integration scheme of deconvolution-clustering-dimension reduction-cross-slice alignment in a unified "cell type composition" low-dimensional semantic space. The basic idea is: modeling the gene expression data by a topic model to obtain a low-dimensional interpretable core representation with cell type composition as the core, directly performing spatial domain division and cross-slice distribution registration on the unified embedding, and realizing spatial consistency through graph constraints, while using a linearly expandable calculation strategy to adapt to multi-slice large-scale data. The present application has the following overall technical effects: 1. Eliminate semantic drift caused by step-by-step processes, so that the deconvolution result and domain division and alignment are optimized in the same representation; 2. The domain boundary is highly consistent with the change of cell composition, which is more consistent with the histological structure; 3. Cross-slice / cross-platform integration is more robust, reducing the risk of over-correction caused by batch differences; 4. The complexity is nearly linearly related to the data size, significantly reducing the time and memory overhead; 5. The output is a unified and interpretable embedding and cell proportion, providing a stable foundation for downstream analysis and visualization.

[0028] The core difference of the present application is that the deconvolution, spatial domain division and multi-slice alignment are completed in the same generation-inference framework based on the cell type composition semantic space as a unified carrier, rather than being processed step by step and disconnected in the expression space as in the existing process. For this purpose, the present application introduces a hierarchical topic model at the UMI level and modulates topic selection with "domain variables": the gene expression profile corresponding to the shared cell type (topic) provides consistent biological semantics across slices, and the topic assignment of UMI is generated by a neural network through a spatial domain indicator vector, so that the cell type distribution is directly coupled with the macroscopic tissue domain in the generation process. The low-dimensional embedding obtained is both a necessary and sufficient statistic for deconvolution, and also a workspace for domain division and alignment, avoiding the semantic drift caused by the three sets of representations of traditional methods "deconvolution-clustering-alignment". This design directly brings two important effects: first, the domain boundary naturally forms with the change of cell type mixture, which is highly interpretable and consistent with histology; second, cross-slice alignment is performed in an interpretable low-dimensional space, which is more robust to batch / platform differences and has lower over-correction risk.

[0029] To achieve computational efficiency and stability on large-scale multi-slices, two key inference and regularization mechanisms are proposed. One is the black-box variational inference of mini-genes: using the separable structure of gene dimension by ELBO, only part of genes are sampled to participate in gradient estimation at each step, reducing time and memory overhead while maintaining unbiasedness, so that end-to-end training of full transcriptome and more than ten slices can be completed on a single card. The second is the domain posterior + graph Laplacian regularization of attention modeling: the neighbor weight integrates the spatial Euclidean distance and the cosine similarity of the posterior distribution, which not only guarantees the spatial continuity of the domain graph, but also maintains sharpness at the real boundary; the Laplacian term further suppresses noise, so that the structure of "smooth within the domain and separate between domains" is still robust at low sequencing depth.

[0030] The WGAN distribution alignment is also introduced in the composition embedding space: the embedding distribution of a certain reference slice is taken as the "true" distribution, and the other slices minimize the Wasserstein distance through the generator and the shared discriminator, and the "domain preservation" constraint can be added, so that the alignment process eliminates batch effects while preserving biological differences and domain structure. Compared with the direct registration in the expression space, this strategy is semantically consistent with the downstream task (domain and composition), and significantly reduces the biological signal flattening caused by over-correction. By combining the above technical features, the application realizes the dual goals of nearly linear computational complexity and biological semantic consistency: the output domain graph and cell proportion are consistent in the same latent space, the cross-slice alignment is robust, and the end-to-end analysis of large-scale multi-slices can be completed on a single machine and a single card.

[0031] The technical solutions of the application will be described in detail below with specific examples. The following specific examples can be combined with each other, and the same or similar concepts or processes may not be described in detail in some examples.

[0032] The method of joint deconvolution domain identification and alignment described in the preferred embodiment of the application, as shown in Figure 1 The method of joint deconvolution domain identification and alignment includes the following steps: In step S101, multi-slice gene transcriptome data and reference data are obtained.

[0033] In one possible implementation, the multi-slice gene transcriptome data includes gene expression count matrix data and spatial coordinate data of multiple slices, and the reference data is single-cell transcriptome data.

[0034] It should be noted that all slices are modeled as a data set in the application, and cell type deconvolution, spatial domain (tissue structure) division, and cross-slice batch alignment are completed simultaneously in the same latent space. Specifically, first, input and preprocessing: receiving gene expression count matrix , spatial coordinates and optional scRNA-seq reference (containing T cell types); then, spatial neighborhood graph construction: construct a sparse graph of neighbors for each slice m with adjacency matrix ; spatial-aware hierarchical topic modeling: treat each UMI read as generated from a "cell type topic" and coupled by a "domain label" to organize structure; variational inference with mini-gene training: estimate posteriors with black-box variational inference and compute gradients in mini-batches per gene; composition representation and deconvolution: aggregate UMI-level topic probabilities at site level into interpretable cell composition representation with softmax to get cell type proportions ; domain clustering and spatial harmonization: cluster domains in embedding space and maintain spatial continuity with graph Laplacian and sparse attention; cross-slice distribution alignment (WGAN): align slices in composition embedding space to remove batch effects.

[0035] In step S102, a variational distribution is obtained according to the multi-slice gene transcriptome data.

[0036] In one possible implementation, a hierarchical topic model is constructed according to the gene expression count matrix data; and the hierarchical topic model is inferred to obtain the variational distribution.

[0037] It is worth noting that, in the embodiments of the present application, a hierarchical topic model is constructed based on the count matrix, and joint modeling and optimization are performed in combination with the spatial coordinates and the reference data, to simultaneously obtain a common latent space of cell type composition, spatial domain probability and cross-slice alignment.

[0038] Specifically, spatial-aware hierarchical topic modeling is performed, assuming that the SRT data (multi-slice gene transcriptome data) contains slices, genes, topics (corresponding to cell types), domains. The th UMI at the th site of each slice d is generated by the following model (hierarchical topic model): ; ; ; ; ; ; wherein, is the gene expression profile of a cell type / topic; read first abstract topic , then abstract topic generate UMI indicator vector ; through neural network with UMI noise embedding with site domain indicator for input, cell type (topic) is used in connection with macro-structure (spatial domain).

[0039] It can be understood that, for slice k , site i UMI read observation; for cell type t gene probability vector at site i ; for Dirichlet distribution prior parameter, control initial gene probability distribution; for topic selection probability at site i , calculated by neural network ; for spatial domain indicator variable (category distribution) at site i ; for UMI noise embedding, subject to normal distribution (the noise parameter has little effect on the whole, and is set as a known parameter in the present application); denotes the product of the covariance matrix as a scalar α and the unit matrix I; " " indicates that the parameter on the left is determined from the data on the right. That is, the present application has observed UMI read observations, and sequentially infers the spatial domain label.

[0040] Specifically, in the process of variational inference and spatial attention, the inference of the posterior distribution of the variables in the above hierarchical topic model adopts black box variational inference (Black Box Variational Inference), and the variational distribution is factorized according to the variable factor, and the expression of the variational distribution is: ; where the posterior probability of each topic is , and the posterior probability of the spatial domain is constructed using the attention mechanism: ; .

[0041] where the composite distance D combines the Euclidean distance and the cosine similarity of the posterior distribution. Thus .

[0042] It can be understood that, is a topic posterior probability, is a domain indicator posterior probability; is a spatial neighborhood attention weight; is a neighborhood weight; is a hyperparameter controlling weight decay; is a neighborhood weight; is a number of spatial domains.

[0043] In step S103, the variational distribution is jointly optimized according to the multi-slice gene transcriptome data and the reference data, to obtain a target parameter.

[0044] In a possible implementation, according to the variational distribution, an evidence lower bound is obtained; the evidence lower bound is optimized according to the spatial coordinate data and the reference data, to obtain a target parameter.

[0045] It should be noted that, is a target parameter to be calculated and determined. is a target parameter to be calculated and determined.

[0046] In a possible implementation, the evidence lower bound is optimized to obtain estimated data; according to the estimated data, a composition representation is obtained; a neighborhood graph is constructed according to the spatial coordinate data, and a target parameter is obtained according to the neighborhood graph, the single-cell transcriptome data and the composition representation.

[0047] It should be noted that, the hierarchical topic model is subjected to black-box variational inference, a variational factorized distribution of latent variables is constructed, including a topic distribution and a spatial domain indicator distribution; a posterior expectation and an evidence lower bound are calculated based on the variational distribution; the variational parameters and the model parameters are jointly updated according to the spatial coordinate and the reference data, to form a joint optimization process.

[0048] After the posterior distribution is constructed, an estimated value of the posterior probability is obtained by optimizing the following evidence lower bound (ELBO), that is, the expression of the evidence lower bound optimization is: ; The obtained site-level composition embedding and cell proportion are: ; ; And let , . Let Collect gene probabilities of all topics (obtained from single-cell reference data).

[0049] It can be understood that, to optimize the objective of variational inference (by a neighborhood graph of the spatial coordinates); to the cell type t the expression probability of the lower gene p . to the unified representation, according to the proportion of the cell type will finally be taken as the composition representation .

[0050] It is worth noting that the evidence lower bound is maximized to obtain the estimation of the topic-related parameters; according to the estimation, a site-level composition representation is obtained, which is a cell type composition vector obtained by aggregating and normalizing the topic variational probability parameters at the gene level according to the gene count; the estimation of the spatial domain indication probability parameter of each site is obtained; a neighborhood graph is constructed according to the spatial coordinate data, and the domain distribution and alignment mapping are updated according to the neighborhood graph, the reference data and the composition representation, so as to complete the joint optimization.

[0051] In a possible implementation, a global optimization objective is obtained according to the neighborhood graph, the single-cell transcriptome data and the composition representation; and a target parameter is determined according to the global optimization objective.

[0052] Specifically, the expression of the global optimization objective is: ; Among them, the third term is a graph Laplace regularization, which promotes the consistency of the low-dimensional representation of the spatial adjacent sites.

[0053] It can be understood that, is the gene expression matrix of the slice k ; is the cell type reference matrix (i.e. single-cell transcriptome data), which stores the topic gene probability; is a graph Laplace regularization term that enforces spatial consistency; is a weight coefficient. The target parameter is determined by global optimization and , so as to determine the joint optimization model.

[0054] It is worth noting that the hierarchical topic model corresponds to the evidence lower bound, the Laplace smoothing term defined on the neighborhood graph is used to encourage spatial continuity; the composition representation regularization term based on the reference data; the target parameter that minimizes the global objective is determined, and the target parameter is fed back to update the variational factorization distribution and the model parameter.

[0055] In a possible implementation, a cell type composition vector is obtained according to the estimation data; a to-be-aligned representation is obtained according to the cell type composition vector, and the to-be-aligned representation is cross-slice aligned to obtain a composition representation.

[0056] Specifically, in the cross-slice alignment process, each slice is treated as a distribution . A reference slice is selected as the "true distribution" (anchor slice) . For each , a generator is trained with a single discriminator , minimizing the Wasserstein distance (with gradient penalty) to obtain , so that is aligned with . A "domain preservation" loss ( for a domain classifier) can be added to suppress over-correction. Alignment is performed in the composition embedding space, and linear correlation is calculated with small batch sampling.

[0057] It can be understood that, after obtaining the unified representation , cross-slice alignment is performed, so as to ensure the accuracy of the composition representation .

[0058] It is worth noting that, the estimation of the probability parameter of the space domain of each site is obtained; the gene count of each site is weighted and aggregated according to the corresponding topic probability to obtain a unified embedding, and a cell type composition vector is obtained through normalization; cross-slice alignment is performed on the composition representation of each slice, the alignment maps the embedding of each slice to the latent space of the reference slice through a learned slice-specific mapping, and is optimized based on a distribution difference measure, so as to obtain the aligned composition representation.

[0059] In one possible implementation, an initial model is constructed, training data is obtained, and the initial model is trained according to the training data to obtain a trained initial model; the multi-slice gene transcriptome data is input into the trained initial model to generate an initial parameter of the variational distribution.

[0060] Specifically, in the model training process, for the alternating optimization of the above model, first enter the topic optimization phase, and maximize the ELBO to update while fixing ; then enter the space domain optimization phase, and update using attention and Laplace regularization while fixing Two phases of cyclic iterations are performed until the ELBO and Laplacian term are stable (about 50 epochs). Meanwhile, a mini-gene strategy is adopted, only sampling ≪P genes to participate in gradient calculation each time, and the unbiased gradient is ensured by the property of ELBO on gene separability in the model assumption, thereby significantly reducing memory and computing time. Spatial message passing relies on sparse adjacency: attention and Laplacian update are performed on the m-nearest neighbor graph, and the complexity is reduced.

[0061] Before inputting the multi-slice gene transcriptome data into the joint optimization process, the model is trained on training data to obtain converged variational parameters and model parameters; then the parameters obtained by training are used as initial parameters for inference and alignment of the data to be analyzed.

[0062] In step S104, according to the target parameters, a deconvolution result, a spatial domain recognition result and unified representation information are generated.

[0063] It can be understood that after obtaining the target parameters, the results are generated according to the target parameters, the unified representation corresponding to the multi-slice gene transcriptome data, the deconvolution result corresponding to the multi-slice gene transcriptome data, the spatial domain recognition result corresponding to the multi-slice gene transcriptome data.

[0064] Referring to Figure 2 , input (left in the figure): spatial transcriptome gene count matrix and its coordinate information from multiple tissue slices, and single-cell transcriptome in the lower left corner as reference data for defining gene expression characteristics of each cell type.

[0065] Referring to Figure 2 , core modeling (upper middle in the figure): fine-grained modeling of each sequencing read (UMI), regarding each sequencing read as coming from a potential "cell type topic", and each topic corresponds to a "topic-specific expression profile". Each spatial site also has a potential "spatial domain" label, and the domain label is involved in the sampling of the "cell type topic" through a neural network. Therefore, the model estimates two types of hidden variables in the same framework: one describes the composition of the site mixed by different cell types (for deconvolution), and the other describes the tissue domain that can be shared across slices (for spatial domain recognition).

[0066] Referring to Figure 2At the site level (lower middle and lower left in the figure), the model aggregates the probability vectors of the "cell type topics" read from the UMI to form the "cell composition representation"; after mapping and normalization, the "cell type proportions" are obtained, and combined with the "cell type-specific gene expression matrix" of the reference data to reconstruct the spatial transcriptome gene count matrix, and the reconstruction error is calculated as the loss, so that the cell composition representation remains interpretable and consistent with the reference.

[0067] Referring to Figure 2 , the method output (right in the figure) is as follows: first, the "cell composition representation" of each spatial site is obtained, and then the cell type proportions are obtained after mapping and normalization, which is the deconvolution result of multiple slices; second, the domain recognition of the site is performed in the same representation space. The system first constructs a sparse spatial neighborhood graph for each slice based on the coordinates, representing the neighborhood relationship of each site, and aggregates information within the local neighborhood using a sparse attention mechanism (upper left in the middle), so that consistent spatial domain division is obtained. Since the domain judgment is directly established on the representation of "cell composition", the domain boundary reflects the change of cell type mixing, rather than the simple fluctuation of gene variance.

[0068] Referring to Figure 2 , cross-slice batch alignment (lower right in the figure): the present application performs distribution matching in the interpretable space of "cell composition representation". The method is to select a reference slice, and the representations of the remaining slices are aligned to the distribution of the reference through light mapping, so as to eliminate batch / platform differences while retaining real biological differences and identified tissue domain structures. The visualization in the lower right corner shows that the slices before and after alignment are well mixed in the same figure, while different tissue domains can still be separated.

[0069] In summary, the outputs of the present application include the cell type proportions of each spatial site (deconvolution), the spatial domain division of each slice, and the aligned unified representation for visualization and subsequent analysis. The key of the whole process is to complete the three tasks in the same "cell composition" representation space, and combine the small batch and sparse attention in the gene dimension: only a small part of genes is processed each time, and information is only propagated between the neighbors, so as to control the memory and time overhead, suitable for multiple slices and full transcriptome scale.

[0070] The present application adopts the mechanism of UMI-level hierarchical topic and domain variable coupling at the generation level: taking as a shared cell type gene probability dictionary, the topic selection of each UMI is determined by the output of two-layer neural network , wherein the domain indication Macroscopic organizational structure is directly incorporated into the topic generation process. As a result, cellular composition and tissue spatial domain structure are unified into the same probabilistic generation model. Domain boundaries are naturally driven by changes in cell type mixing, rather than solely by heterogeneity in gene expression, thus significantly enhancing interpretability and histological consistency.

[0071] To obtain a continuous and clear spatial domain graph, this application constructs a domain posterior through an attention mechanism in variational inference: adjacency weights. Simultaneously relying on spatial Euclidean distance and cosine similarity of the posterior distribution, via Output domain probability This composite metric enables neighboring regions to support each other probabilistically to maintain spatial continuity, while posterior similarity sharpens the true anatomical boundaries; thus, stable and noise-robust domain partitioning can still be obtained in slices with irregular topology or local density heterogeneity.

[0072] Regarding inference efficiency, this application proposes a black-box variational inference algorithm for mini-genes: sampling only once per iteration. Each gene participates in gradient calculation. ELBO is used to ensure unbiased gradient estimation based on the separable structure of the genes, thus reducing time and memory overhead. linearly reduced to This feature directly brings considerable engineering benefits: a dataset with the entire transcriptome and more than ten slices can be trained on a single general-purpose GPU, significantly improving throughput and scalability.

[0073] To enhance spatial consistency and suppress noise, this application incorporates compositional characterization. Applying a graph Laplace regularization This approach promotes convergence in the characterization of adjacent sites within the same domain, while preserving the characterization intervals across spatial domains, thus achieving the effect of "smooth intra-domain and separated inter-domain"; the domain structure remains robust even at low sequencing depths or in cases of local site loss.

[0074] Regarding cross-slice integration, this invention performs GAN distribution alignment in an interpretable compositional representation space rather than the original expression space. Since the input is a low-dimensional representation of cell type composition, the generation of a discriminative game tends to correct for distribution drift caused by slices / platforms while preserving biological differences to the greatest extent possible. Compared to direct registration in the expression space, this significantly reduces the risk of overcorrection and improves consistency across platforms and individuals.

[0075] The proposed solution was validated on multiple public datasets. A systematic comparison was conducted between this invention (SALID) and eight mainstream algorithms on 12 10×Visium human dorsolateral prefrontal cortex (DLPFC) slices (see [link to relevant documentation]). Figure 3a). The spatial domains reconstructed by the present application clearly delineate the laminar structure of L1-L6 and white matter: oligodendrocytes are highly abundant in white matter and layer 6, while astrocytes are enriched in superficial L1 / L2, which are highly consistent with histological annotations; while BayesSpace and other methods show ambiguous boundaries at the junction of layers 2-3, and SEDR and STAGATE even misalign white matter with superficial layers. Quantitatively, the spatial domains of the present application have a matching degree (ARI) of 0.64±0.03 with artificial annotations, which is better than all reference tools in three groups of sections (see Figure 3 b). The WGAN alignment strategy of the present application effectively eliminates batch effects between sections: in the UMAP space, the same layer domains of different sections are well mixed, while PCA and other algorithms still show obvious section clustering. The iLIST score shows that the cross-section embedding consistency of the present application is improved by about 20%, and the best batch correction effect is achieved while maintaining the layer domain resolution (ARI=0.58) (see Figure 3 c).

[0076] The present application is further applied to high-dimensional Stereo-seq mouse embryo time series (E9.5-E16.5, a total of 8 sagittal sections, a total of about 540,000 spatial sites). Figure 4 a shows the comparison between the spatial domains recognized by the present application and the reference embryo dissection annotations: in all development stages, the present application clearly separates anatomical structures such as heart, liver bud, brain-spinal cord tube, meninges, and primitive endoderm, and the boundaries are highly consistent with the published embryo dissection map. In order to quantitatively track the organ embryogenesis dynamics, we selected four typical regions: liver, brain and spinal cord, meninges, and heart, and counted the cell composition abundance changes in the 8 sections (see Figure 4 b). The region volume proportions output by the present application are highly consistent with known embryo morphology: the liver bud increases from <3% at E9.5 to >20% at E16.5, the heart volume reaches a peak at E11.5 and then stabilizes, and the meninges and brain-spinal cord tube show a continuous expansion-differentiation trend; these trends are consistent with histological observations. In the spatial quantitative evaluation (see Figure 4 c), the Pearson correlation coefficients of the present application for the liver, heart, meninges, brain-spinal cord, and lung (formed later) five anatomical domains reach an average of 0.85. The overall results show that the present application not only maintains high accuracy in static adult tissues, but also reliably analyzes rapidly changing embryonic morphology, laying a data-driven foundation for subsequent three-dimensional multi-section analysis.

[0077] In summary, the present application realizes near-linear computation expansion while ensuring biological interpretability. The resulting domain map is consistent with the cell proportion in the same latent semantic space, robustly aligned across slices, and can provide end-to-end deconvolution, domain detection, and batch correction results on large-scale multi-slice data.

[0078] In terms of prior and theme dictionary setting, the present application is not limited to using Dirichlet prior; it can also be replaced by Logit-Normal and other continuous priors, and on this basis, group sparsity or hierarchical structure can be introduced to express the subordinate and shared relationship between cell lineages. The acquisition of the cell type reference matrix can also be directly learned during training, or initialized using scRNA-seq reference and kept consistent through a penalty loss, or first obtain an interpretable initial dictionary through unsupervised decomposition such as matrix non-negative binomial decomposition, and then fine-tune in the generative model.

[0079] In terms of variational inference and optimization, the variational distribution can be flexibly adjusted according to data and computing power. The UMI noise embedding can use a reparameterizable Gaussian family; for discrete domain variables the Concrete / Gumbel-Softmax approximation can be used for backpropagation. The optimizer is not limited to Adam, but can also use AdamW or Lookahead; and can combine temperature annealing, early stopping, gradient clipping, mixed precision, and distributed data parallel training techniques to obtain a more stable and efficient convergence process.

[0080] The spatial graph and message passing mechanism also have multiple interchangeable implementations. Spatial adjacency can be constructed using a radius graph or a hexagonal grid, and the number of neighbors can be adaptive or use a multi-scale strategy to balance local and global structures; the composite metric D can be given by the weighted sum of Euclidean distance and posterior cosine similarity, or can be replaced by a learnable metric network to adapt to complex organizational geometry. In addition to the current attention implementation, message passing can also use GAT, GraphSAGE, MRF / CRF, and other equivalent ways; and the Laplace regularization used for spatial consistency can also be replaced by total variation or anisotropic smoothing to better preserve sharp boundaries.

[0081] The cross-slice alignment module is not limited to WGAN. In the composition embedding space, maximum mean difference matching or an optimal transport (OT) method can also be used, and the above methods can be combined as needed to improve robustness. The reference distribution can be fixed to a certain reference slice, or the most matching slice can be dynamically selected, or the Wasserstein barycenter can be calculated as a global common reference; if stronger structure preservation is required during alignment, a discriminator conditioned on "domain" or "class" can be introduced (conditional alignment), which eliminates batches while maintaining biological stratification.

[0082] The specific implementation of the composition embedding and domain partitioning also allows for variations. The number of domains L can be pre-set or adaptively determined by profile coefficients, BIC, or sparse Dirichlet process criteria. In terms of clustering algorithms, a discriminative classifier based on Softmax can be used, or Kmeans or spectral clustering can be used to obtain more suitable segmentation patterns in different tissue scenarios.

[0083] Regarding the reference usage mode, the application simultaneously supports three paradigms: supervised, semi-supervised, and unsupervised. When a scRNA-seq reference is available, the theme is aligned with the reference expression by a penalty loss; when there is no reference, the generated model and spatial regularization are independently driven; a small number of labeled genes can also be used as weak supervision signals to improve the interpretability and stability of deconvolution and domain recognition.

[0084] At the engineering and deployment level, the application can run in a lightweight environment with a single machine and a single card (≥12 GB of video memory), or it can be extended to multiple cards or clusters; command line tools and Python APIs are provided to facilitate batch processing and pipeline integration, and Docker containers can be used for out-of-box replication and cross-platform deployment. To further improve large-scale training efficiency, mini-gene sampling can use importance sampling strategies (weighted by gene variance or abundance), combined with gradient accumulation to handle larger batches and more cores under limited video memory.

[0085] It should be noted that the "cell composition embedding" is generated using matrix factorization or variational autoencoder (VAE), and then the domain and alignment are performed on the embedding; or a discriminative network (GAN) is used to end-to-end regress cell proportions and jointly predict domain labels, and an additional distribution registration module is added to replace the application embodiment scheme as a whole. The above alternatives differ in implementation path, but if they are unified in the "cell composition representation" as the only working space to perform the three tasks, they still belong to the technical idea of the application in essence.

[0086] Around the deconvolution and topic model prior, Logit-Normal, sparse Gaussian prior, hierarchical prior or direct learnable dictionary can be used instead; UMI-level modeling can also be replaced by a polynomial / Poisson decomposition of the total count per site to obtain cell proportions. Such transformations change the prior and granularity, but as long as the final site-level proportion / composition embedding is used as the common space for domain and alignment, the technical effect is equivalent.

[0087] In terms of domain posterior and spatial consistency, "attention mechanism + Laplace" can be replaced by MRF potential function or message passing of GNN (GraphSAGE / GAT), or even only morphological operation or superpixel map for spatial smoothing. They are different in algorithm form, but the common goal is to propagate / constrain the composition embedding within the neighborhood, so that the domain map is both continuous and can preserve the boundary.

[0088] In terms of cross-slice alignment, instead of using adversarial (WGAN), non-adversarial distribution registration such as optimal transport (OT) or Gaussian matching can be used; a "domain adaptive" discriminator can also be used without using Wasserstein distance. As long as the alignment occurs in the composition representation space rather than the original expression space, the technical effect is still "batch correction while preserving biological differences".

[0089] In terms of representation and domain division, the obtained low-dimensional representation can be further processed through a contrastive learning head to generate another representation and then clustered, or the Softmax proportion can be replaced by a Logistic-Normal proportion; or the Kmeans / Softmax classifier can be replaced by spectral clustering / Dirichlet process, or the number of domains L can be adaptively determined by information criterion / sparse prior. These will not change the key effect of "completing domain identification in the composition semantic space".

[0090] It is worth noting that "unified cell composition semantic space" can be regarded as a transferable modeling principle rather than a one-time engineering implementation. This space couples the objective functions of deconvolution, domain division, and cross-slice alignment in a set of inference closed loop, making the relationship between "domain boundary changes with composition" an endogenous property of the model. Subsequent analysis of identifiable and consistent can be given under this principle. In the engineering and landing level, the sparse graph design of the application is naturally suitable for lightweight and heterogeneous hardware acceleration. In practice, mixed precision and parameter grouping update can be used to further reduce memory overhead; in the environment of limited computing power or data island, the inference can be divided into two-stage process, the former is completed on the edge node, and the latter only transmits low-dimensional composition embedding and statistics, so as to balance efficiency and privacy. Secondly, the system for joint deconvolution domain identification and alignment according to the embodiments of the application is described with reference to the accompanying drawings, which is applied to the method for joint deconvolution domain identification and alignment in any of the above-mentioned solutions.

[0091] Figure 5 Figure 1 is a structural diagram of a system for joint deconvolution domain identification and alignment according to an embodiment of the present application.

[0092] As shown in Figure 5 the joint deconvolution domain identification and alignment system comprises a data acquisition module 100, a variational inference module 200, a joint optimization module 300 and a result output module 400.

[0093] Specifically, the data acquisition module 100 is configured to acquire multi-slice gene transcriptome data and reference data. The variational inference module 200 is configured to obtain a variational distribution according to the multi-slice gene transcriptome data. The joint optimization module 300 is configured to jointly optimize the variational distribution according to the multi-slice gene transcriptome data and the reference data, to obtain a target parameter. The result output module 400 is configured to generate a deconvolution result, a spatial domain identification result and unified representation information according to the target parameter.

[0094] Figure 6 Figure 2 is a structural diagram of a terminal according to an embodiment of the present application. The terminal can comprise: a memory 501, a processor 502 and a computer program stored in the memory 501 and executable on the processor 502.

[0095] The processor 502 implements the method for joint deconvolution domain identification and alignment provided in the above embodiments when executing the program.

[0096] Further, the terminal further comprises: a communication interface 503 for communication between the memory 501 and the processor 502.

[0097] The memory 501 is configured to store the computer program executable on the processor 502.

[0098] The memory 501 can include a high-speed RAM memory, and can also include a non-volatile memory such as at least one disk memory.

[0099] 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 connected with each other through a bus and complete communication between each other. The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component (PCI) bus or an Extended Industry Standard Architecture (EISA) bus, etc. The bus can be divided into an address bus, a data bus, a control bus, etc. For convenience of representation, Figure 6 Only one thick line is used to represent the bus in the figure, but it does not mean that there is only one bus or only one type of bus.

[0100] Optionally, in a specific implementation, if the memory 501, the processor 502 and the communication interface 503 are integrated on a chip, the memory 501, the processor 502 and the communication interface 503 can complete communication between each other through an internal interface.

[0101] The processor 502 can be a Central Processing Unit (CPU), or an Application Specific Integrated Circuit (ASIC), or one or more integrated circuits configured to implement one or more embodiments of the present application.

[0102] The embodiment also provides a computer readable storage medium, which stores a computer program, and the program is executed by a processor to implement the method for joint deconvolution field identification and alignment.

[0103] An embodiment of the present application provides a computer program product, which comprises a computer program, and the computer program is executed by a processor to implement the method for joint deconvolution field identification and alignment. Figure 1 The corresponding embodiment provides the method for joint deconvolution field identification and alignment.

[0104] In the description of the application, reference to "one embodiment", "some embodiments", "an example", "a specific example", or "some examples" means that a particular feature, structure, material, or characteristic being described is included in at least one embodiment or example of the application. The appearances of the phrase in various places in the specification are not necessarily all referring to the same embodiment or example. Furthermore, the described specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples. In addition, the usage of "N" means at least two, for example, two, three or the like, unless explicitly stated otherwise.

[0105] Furthermore, the terms "first", "second", or the like, are used merely as a designation of certain elements or features, and do not imply or connote relative importance or a specific order of categorization of the indicated features. Accordingly, features described as "first" or "second" can be explicitly or implicitly included in at least one of the features. In the description of the application, the meaning of "N" is at least two, for example, two, three, etc., unless explicitly specified otherwise.

[0106] Any process or method descriptions or blocks in flow charts or otherwise described herein represent embodiments which can be managed as one or more modules, segments, or portions of code which include one or more executable instructions for implementing specific logic functions or steps, and alternate implementations are possible. In some embodiments, the processes and methods described can be executed by one or more apparatuses or devices, either directly or after conversion to another language. Alternate implementations are possible.

[0107] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable storage medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-included system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable storage medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device. More specific examples (a non-exhaustive list) of computer-readable storage media include: an electrical connection having one or more wires (electronic device), a portable computer disk drive (magnetic device), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Alternatively, the computer-readable storage medium could be paper or other suitable media on which the program can be printed, since the program can be obtained electronically by optically scanning the paper or other medium, followed by editing, interpreting, or otherwise processing as necessary, and then stored in a computer memory.

[0108] It should be understood that the various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, the N steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0109] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0110] In addition, each of the function units in each of the embodiments of the present application can be integrated in one processing module, or each unit can be physically present separately, or two or more units can be integrated in one module. The integrated module can be realized in the form of hardware or in the form of a software function module. When the integrated module is realized in the form of a software function module and sold or used as an independent product, it can also be stored in a computer readable storage medium.

[0111] The storage medium mentioned above can be a read-only memory, a magnetic disk or an optical disk, etc. Although the embodiments of the present application have been shown and described above, it should be understood that the above embodiments are exemplary and should not be construed as limiting the present application, and those skilled in the art can make changes, modifications, replacements and variations to the above embodiments within the scope of the present application.

[0112] It should be understood that the application of the present application is not limited to the above examples, and those skilled in the art can improve or change the above description, and all these improvements and changes should belong to the protection scope of the claims of the present application.

[0113] Finally, it should be pointed out that: the above embodiments are only used to illustrate the technical solutions of the present application, and not to limit them; although the present application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that: it can still modify the technical solutions recorded in the foregoing embodiments, or make equivalent replacement for part or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. A method for joint deconvolution domain recognition and alignment, characterized in that, The method for joint deconvolutional domain recognition and alignment includes: Obtain multi-slice genome transcriptome data and reference data; Based on the multi-slice gene transcriptome data, the variational distribution was obtained; The variational distribution is jointly optimized based on the multi-slice gene transcriptome data and the reference data to obtain the target parameters; Based on the target parameters, deconvolution results, spatial domain recognition results, and unified representation information are generated.

2. The method for joint deconvolution domain identification and alignment according to claim 1, characterized in that, The multi-slice transcriptome data includes gene expression count matrix data from multiple slices; The step of obtaining the variational distribution based on the multi-slice transcriptome data specifically includes: A hierarchical topic model was constructed based on the gene expression count matrix data; The variational distribution is obtained by inferring the hierarchical topic model.

3. The method for joint deconvolution region identification and alignment according to claim 2, characterized in that, The multi-slice transcriptome data also includes spatial coordinate data of multiple slices; The joint optimization of the variational distribution based on the multi-slice transcriptome data and the reference data to obtain the target parameters specifically includes: Based on the variational distribution, the lower bound of the evidence is obtained; The lower bound of the evidence is optimized based on the spatial coordinate data and the reference data to obtain the target parameters.

4. The method for joint deconvolution region identification and alignment according to claim 3, characterized in that, The reference data is single-cell transcriptome data; The optimization of the lower bound of evidence based on the spatial coordinate data and the single-cell transcriptome data to obtain the target parameters specifically includes: The lower bound of the evidence is optimized to obtain estimated data; Based on the estimated data, the compositional characterization is obtained; A neighborhood graph is constructed based on the spatial coordinate data, and the target parameters are obtained based on the neighborhood graph, the single-cell transcriptome data, and the compositional characterization.

5. The method for joint deconvolution region identification and alignment according to claim 4, characterized in that, The step of constructing a neighborhood graph based on the spatial coordinate data, and obtaining target parameters based on the neighborhood graph, the single-cell transcriptome data, and the compositional characterization, specifically includes: Based on the neighborhood graph, the single-cell transcriptome data, and the compositional characterization, a global optimization objective is obtained. The target parameters are determined based on the global optimization objective.

6. The method for joint deconvolution region identification and alignment according to claim 5, characterized in that, The step of obtaining the compositional characterization based on the estimated data specifically includes: The cell type composition vector is obtained based on the estimated data; The compositional characterization is obtained by aligning the cell type compositional vector and then performing cross-slice alignment on the characterization to obtain the compositional characterization.

7. The method for joint deconvolution domain identification and alignment according to claim 3, characterized in that, The step of jointly optimizing the variational distribution based on the multi-slice transcriptome data and the reference data to obtain the target parameters also includes: Build an initial model, obtain training data, and train the initial model based on the training data to obtain a trained initial model; The multi-slice gene transcriptome data is input into the trained initial model to generate the initial parameters of the variational distribution.

8. A system for joint deconvolution domain recognition and alignment, characterized in that, The system for joint deconvolution domain identification and alignment is applied to the method for joint deconvolution domain identification and alignment as described in any one of claims 1-7; The system for joint deconvolutional domain recognition and alignment includes: The data acquisition module is used to acquire multi-slice gene transcriptome data and reference data; The variational inference module is used to obtain the variational distribution based on the multi-slice gene transcriptome data; The joint optimization module is used to jointly optimize the variational distribution based on the multi-slice gene transcriptome data and the reference data to obtain target parameters. The result output module is used to generate deconvolution results, spatial domain recognition results, and unified representation information based on the target parameters.

9. A terminal, characterized in that, The terminal includes: a memory, a processor, and a program for joint deconvolution domain identification and alignment stored in the memory and executable on the processor. When the program for joint deconvolution domain identification and alignment is executed by the processor, it implements the steps of the method for joint deconvolution domain identification and alignment as described in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a program for joint deconvolution domain identification and alignment, which, when executed by a processor, implements the steps of the method for joint deconvolution domain identification and alignment as described in any one of claims 1-7.

Images