A dual-track parallel purification method, apparatus, medium, and program product for high-dimensional spatial transcriptomic data

Abstract

The embodiment of the specification provides a double-track parallel purification method, equipment, medium and program product for high-dimensional space transcriptome data, and relates to the field of intelligent medical treatment. The method comprises the following steps: acquiring a virtual full transcriptome expression matrix of an image block in a histopathological full section image of a target disease, wherein the virtual full transcriptome expression matrix is predicted and output by a generative spatial transcriptome prediction model on the image block; screening a target gene expression subset from the virtual full transcriptome expression matrix according to a preset immune microenvironment related gene set; performing double-track parallel dimension reduction processing on the target gene expression subset, and respectively generating a modular feature matrix and a microenvironment latent feature matrix; and merging the modular feature matrix and the microenvironment latent feature matrix to obtain a denoised pure virtual molecular feature matrix. Through the double-track parallel purification method combining biological prior knowledge and unsupervised mining, required molecular features are extracted from massive high-dimensional noise data.

Description

Technical Field

[0001] This invention relates to the field of intelligent healthcare, and more specifically, to a dual-track parallel purification method, apparatus, medium, and program product for high-dimensional spatial transcriptome data. Background Technology

[0002] With the rapid development of digital pathology and spatial transcriptomics, analyzing tumor microenvironment heterogeneity using multimodal data has become an important direction for precision medicine. In recent years, the emergence of generative spatial transcriptomics models has made it possible to predict genome-wide expression levels using routine H&E-stained pathological images, reducing the cost of acquiring spatial molecular information.

[0003] However, in practical applications, the virtual whole transcriptome data directly output by such generative models faces severe problems of "curse of dimensionality" and "noise interference".

[0004] Specifically, the model's predicted output typically covers the expression levels of over 30,000 genes, the vast majority of which have no direct biological connection to specific clinical problems. Furthermore, the prediction process itself introduces a significant amount of non-biological computational noise. Existing dimensionality reduction methods often employ global principal component analysis or a single autoencoder architecture. While these methods can mathematically compress data dimensionality, they often neglect prior biological knowledge of gene expression, resulting in a lack of interpretability of the reduced features and even the loss of crucial immune microenvironment signals.

[0005] Furthermore, due to the lack of a specific purification mechanism for high-dimensional virtual gene data, directly using the original prediction data for downstream multimodal fusion or prognostic modeling can easily lead to model overfitting and fail to accurately capture subtle differences in immune status in complex tumors such as nasopharyngeal carcinoma. Summary of the Invention

[0006] This invention aims to at least solve one of the technical problems existing in the prior art. To this end, this invention provides a dual-track parallel purification method, apparatus, medium, and program product for high-dimensional spatial transcriptome data; the method of this invention, by combining biological prior knowledge with unsupervised mining strategies, extracts low-dimensional, high signal-to-noise ratio molecular features with clear biological significance from massive high-dimensional noisy data.

[0007] The first aspect of this application discloses a dual-track parallel purification method for high-dimensional spatial transcriptome data, comprising: A virtual whole transcriptome expression matrix is ​​obtained from the image patch in the whole tissue pathology image of the target disease. The virtual whole transcriptome expression matrix is ​​predicted and output by the generative spatial transcriptome prediction model for the image patch. Based on a preset set of immune microenvironment-related genes, a subset of target gene expressions is selected from the virtual whole transcriptome expression matrix; The target gene expression subset is subjected to dual-track parallel dimensionality reduction processing: In the first processing track, a core gene subset that meets the preset immune relevance is extracted from the target gene expression subset. The genes in the core gene subset are divided into N functional modules according to their biological functions. The statistical characteristics of gene expression in each functional module are calculated to generate a modular feature matrix. In the second processing track, nonnegative matrix factorization is performed on the expression levels of the remaining genes in the target gene expression subset to extract the basis matrix as the potential feature matrix of the microenvironment. The modular feature matrix and the microenvironment latent feature matrix are combined to obtain the denoised pure virtual molecular feature matrix.

[0008] A second aspect of this application discloses a computer device, comprising: a memory and a processor; the memory is used to store a computer program; the processor executes the computer program to implement the steps of the above-described method.

[0009] A third aspect of this application discloses a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method described above.

[0010] The fourth aspect of this application discloses a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method.

[0011] This application has the following beneficial effects: 1. This application innovatively proposes a "dual-track parallel" dimensionality reduction and purification method, which effectively solves the problem of extremely high dimensionality and huge noise in generative spatial transcriptome data.

[0012] By using the first processing track and leveraging prior knowledge of the nasopharyngeal carcinoma immune microenvironment, core immune-related genes are mapped to specific biological functional modules, and statistical characteristics within the modules are calculated. This not only compresses the data dimensions but also preserves the biological interpretability of key immune pathways.

[0013] Through a second processing track, nonnegative matrix factorization is performed on the remaining genes to automatically uncover unknown potential co-expression patterns. This combination of "prior guidance + unsupervised mining" ensures that the final output of a pure virtual molecular feature matrix has both clear biological meaning and encompasses the deep structural information implicit in the data.

[0014] 2. In this application, the purified virtual molecular features can achieve spatial alignment and vector splicing with the macroscopic morphological features of the same image block at the bottom layer, and the generated multimodal fusion feature matrix effectively breaks the feature space limitation of a single modality.

[0015] Furthermore, by combining central log ratio transformation and regularized Cox survival analysis model, this application can transform complex phenotypic profiles into quantifiable patient prognostic risk scores, achieving automated prediction from pathological images to clinical decision support, and improving the accuracy and robustness of prognostic models.

[0016] In addition, in the division of functional modules and non-negative matrix factorization, the specific number of genes or modules in this application is a dynamic adjustment mode designed to adapt to different analysis needs. This setting makes this method not only suitable for nasopharyngeal carcinoma, but also applicable to other tumor types with similar high-dimensional noise problems. Attached Figure Description

[0017] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a schematic diagram of the method flow provided in the first aspect of the present invention; Figure 2 This is a schematic diagram of a dual-track parallel purification system for high-dimensional spatial transcriptome data provided in an embodiment of the present invention; Figure 3 This is a flowchart of the multimodal feature extraction and splicing method provided in the embodiments of the present invention; Figure 4 This is a schematic diagram of the virtual molecular feature noise reduction principle provided in the embodiments of the present invention; Figure 5 This is a time-series diagram comparing tearing and reconstruction in HPCs based on the Leiden algorithm, provided in an embodiment of the present invention. Figure 6 This is a network topology diagram for predicting the prognostic risk of immunotherapy based on MP-HPCs provided in an embodiment of the present invention. 。 Detailed Implementation

[0019] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0020] In some of the processes described in the specification, claims, and accompanying drawings of this invention, multiple operations appearing in a specific order are included. However, it should be clearly understood that these operations may not be executed in the order they appear herein, or may be executed in parallel. The operation numbers, such as 101, 102, etc., are merely used to distinguish different operations and do not represent any execution order. Furthermore, these processes may include more or fewer operations, and these operations may be executed sequentially or in parallel. It should be noted that the descriptions such as "first," "second," etc., in this document are used to distinguish different messages, devices, modules, etc., and do not represent a sequential order, nor do they limit "first" and "second" to different types.

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] like Figure 1 As shown in the diagram, this embodiment of the invention provides a schematic flowchart of a dual-track parallel purification method for high-dimensional spatial transcriptome data. Specifically, the method includes the following steps: S101: Obtain the virtual whole transcriptome expression matrix of the image patch in the whole tissue pathology image of the target disease. The virtual whole transcriptome expression matrix is ​​predicted and output by the generative spatial transcriptome prediction model for the image patch. In some embodiments, the object of the target disease in this application is a test sample. The terms "subject," "test subject," or "test sample" as used herein refer to any animal (e.g., a mammal), including but not limited to humans, non-human primates, rodents, etc., which will become the recipient of a specific treatment. Generally, the terms "subject" and "patient" are used interchangeably herein when referring to human subjects. Preferably, the subject is a human.

[0023] In this embodiment, the target disease can be nasopharyngeal carcinoma, but it can also be other tumor types that have similar high-dimensional noise problems. The tumor type is not specifically limited in this embodiment.

[0024] In this embodiment, nasopharyngeal carcinoma is used as the target disease to illustrate the source and processing of whole-section histopathological images. However, this description is only one use case and is not intended to limit the scope of protection of this data processing method. Specifically, the exemplary description is as follows: Step 1: Obtain the whole-slide image (WSI) file of nasopharyngeal carcinoma pathology generated by the whole-field digital slide scanner.

[0025] Step 2: The server executes the tissue segmentation and dicing algorithm. At a magnification of 10X or 20X, the effective tissue region of WSI is divided into 224 x 224 pixel non-overlapping image blocks.

[0026] Step 3: Perform color standardization. The server uses the Reinhard algorithm to perform color space conversion on the extracted image patches, eliminating H&E staining color deviations caused by different institutions and scanners, and outputting a standardized set of image patches.

[0027] In this stage, the server mainly completes the physical segmentation and color normalization of the input data, providing a standard input source for subsequent deep learning networks.

[0028] S102: Based on the preset immune microenvironment-related gene set, select the target gene expression subset from the virtual whole transcriptome expression matrix; In some embodiments, the preset immune microenvironment-related gene set includes genes associated with the immunotherapy response to the target disease; In a more specific embodiment, still taking nasopharyngeal carcinoma as an example, it mainly includes step 1 (prior knowledge screening): the server reads a preset gene dictionary of nasopharyngeal carcinoma immune microenvironment characteristics, and... The matrix performs feature index filtering, retaining only 862 gene features that are highly associated with nasopharyngeal carcinoma immunotherapy.

[0029] S103: Perform dual-track parallel dimensionality reduction on the target gene expression subset; In the first processing track, a core gene subset that meets the preset immune relevance is extracted from the target gene expression subset. The genes in the core gene subset are divided into N functional modules according to their biological functions. The statistical characteristics of gene expression in each functional module are calculated to generate a modular feature matrix. In the second processing track, nonnegative matrix factorization is performed on the expression levels of the remaining genes in the target gene expression subset to extract the basis matrix as the potential feature matrix of the microenvironment. In some embodiments, the core gene subset includes M key biomarker genes that conform to a preset immune correlation, where M is a natural number greater than 5.

[0030] In some embodiments, the functional modules are divided according to immune cell subsets or immune pathway functions, where N is a natural number greater than or equal to 2, and include at least two of the following modules: T cell infiltration module, T cell exhaustion module, and macrophage planning module.

[0031] In some embodiments, the statistical features include the mean and / or coefficient of variation calculated after standardizing the expression values ​​of each gene within the functional module.

[0032] In some embodiments, the decomposition rank of the non-negative matrix factorization in the second processing track is a settable parameter, and a preset number of potential gene co-expression patterns are automatically extracted by performing non-negative matrix factorization.

[0033] S104: Merge the modular feature matrix and the microenvironment potential feature matrix to obtain the denoised pure virtual molecular feature matrix.

[0034] In a more specific embodiment, the dual-track parallel dimensionality reduction process specifically involves asynchronously processing the above-mentioned splitting of the 862 genes into two data streams.

[0035] Sub-path A (Functional module mean-based calculation): The server extracts 100 core genes and maps them to 16 functional modules (such as T cell infiltration, stem T cells, TLS chemotaxis and structural modules, etc.) according to biological prior logic. For each module... Gene expression levels within the cell were first normalized using Z-scores. (in Gene expression level, and (This represents the mean and standard deviation of gene j). Then, the mean and coefficient of variation of all genes within this module are calculated. Finally, each module is condensed into two statistical features, and the combined output is... Modular feature matrix of dimension The physical significance of this calculation formula lies in smoothing out the prediction noise of a single gene and improving the noise resistance of key immune pathways.

[0036] Sub-path B (non-negative matrix factorization mining): The server processes the remaining 762 gene expression matrices. Perform an unsupervised nonnegative matrix factorization (NMF) algorithm. The algorithm sets the rank of the factorization. By iteratively optimizing the objective function (W ≥ 0, H ≥ 0), find the basis matrix W and coefficient matrix H, and automatically extract potential and unknown microenvironment gene co-expression patterns. The server extracts matrix W as... 3D microenvironment latent feature matrix Ultimately, the high-dimensional noise of the molecular branch was significantly compressed, resulting in a pure virtual molecular feature output with a dimension of 60 (32+28).

[0037] In some embodiments, the method further includes: Obtain the macroscopic morphological feature matrix of the image patch in the whole tissue pathology slice image described in this application. The macroscopic morphological feature matrix is ​​obtained by extracting features from the same image patch using a visual feature extraction model. The pure virtual molecular feature matrix and the macroscopic morphological feature matrix obtained based on the same image patch are concatenated at the image patch level to generate a multimodal fusion feature matrix; A nearest neighbor graph network is constructed based on the multimodal fusion feature matrix, and clustering is performed using a community detection algorithm to generate multimodal morphological-molecular phenotypic clusters.

[0038] In a more specific embodiment, the sources of the macroscopic morphological feature matrix and the virtual whole transcriptome expression matrix include: Step 1 (Historical Morphology Branch Processing): The server calls a self-supervised visual foundation model (UNI Encoder based on the Vision Transformer architecture) pre-trained on massive amounts of pathological images. A single image patch is input into the model to extract deep network features, and the output dimension is... The initial morphological feature vector. To reduce computational complexity and remove redundant information, the server performs L2 regularization on this 1024-dimensional vector. After standardization (Z-scoreScaler), principal component analysis (PCA) is called to reduce the dimensionality of the feature matrix to [value missing]. Dimension, thus obtaining a refined macroscopic morphological feature matrix. , where N is the total number of image blocks.

[0039] Step 2 (Gene Expression Branching): The server inputs the same image patch into the Generative Spatial Transcriptome Model (STPath). The STPath network learns the mapping relationship between tissue morphology and the spatial transcriptome, predicts the virtual gene expression levels of approximately 38,984 genes within the image patch region, and outputs a virtual whole transcriptome expression matrix. .

[0040] In a more specific embodiment, the server performs low-level fusion of data streams from different modalities and completes the construction of phenotypic maps. The specific process is as follows: Step 1 (Vector Concatenation): The server concatenates the morphological matrix at the image patch level. (128-dimensional) and the denoised molecular feature matrix (32-dimensional) (28-dimensional) Perform a stitching operation to generate Multimodal fusion feature matrix .

[0041] Step 2 (Graph Network Construction and Leiden Clustering): The server calculates the Euclidean distance between each pair of image patches on the 188-dimensional multimodal matrix, constructing a K-Nearest Neighbors (KNN) graph network. Subsequently, the Leiden community detection algorithm is called to perform unsupervised clustering optimization on the network. The internal implementation principle of the Leiden algorithm is to maximize the modularity H of the graph, and its calculation formula is:

[0042] The physical meanings of the formula parameters are as follows: k represents the actual number of edges assigned to cluster (community) c. C Let represent the sum of the degrees of all nodes in cluster c (i.e., the expected number of edges), and m be the total number of edges in the network. γ is the resolution parameter used to control the granularity of clustering. Algorithm execution sequence: First, each node is treated as an independent community. Then, the modularity gain when moving a node to an adjacent community is iteratively evaluated. If the gain is positive, the node is moved until the network reaches a local optimum.

[0043] Step 3 (Clustering Reconstruction Effect): When using only 128-dimensional morphological features, Leiden clustering outputs 35 basic tissue morphological phenotypic clusters (HPCs). This scheme, by injecting 60-dimensional molecular signals, alters the topological connectivity of the KNN graph network. Originally morphologically similar HPC nodes are automatically split and reconstructed under the repulsive force of molecular features, ultimately outputting 53 more detailed morphology-molecular multimodal phenotypic clusters (MP-HPCs).

[0044] In some embodiments, the method further includes: The proportion of image patches for each category of multimodal morphology-molecular phenotype clusters in a single histopathological whole-section image is calculated to obtain the component vector; Perform a central logarithmic ratio transformation on the component vector to obtain the transformed feature vector; The transformed feature vector is input into a pre-trained Cox survival analysis model to obtain a prognostic risk score; The Cox survival analysis model incorporates a regularization penalty mechanism during training. Using the survival time and survival status of patients in the training cohort as target variables, feature selection is performed on the transformed feature vector. The regression coefficients of features that contribute less than a preset threshold to prognostic prediction are compressed to zero, and only features with non-zero regression coefficients and their corresponding weights are retained. The prognostic risk score is obtained by linearly weighting the transformed values ​​of the retained features and their corresponding weights.

[0045] In some more specific embodiments, for end-to-end clinical applications, the server performs survival prediction assessment based on the calculated MP-HPCs atlas, and the specific process is as follows: Step 1: Count the number of image patches belonging to each MP-HPC in a single patient's WSI, divide by the total number of image patches in that WSI, and obtain the MP-HPC spatial proportion component vector with dimension 53. ,in .

[0046] Step 2: To eliminate collinearity of component data and satisfy the independence assumption of Cox regression, the server executes the Center Log-Ratio (CLR) transformation algorithm. The formula principle is as follows:

[0047] in, The geometric mean of the component vectors: .

[0048] Step 3: Input the transformed features into the pre-trained LASSO-Cox survival analysis model. The model calculates the sum of risks based on the predefined regression coefficients of each MP-HPC and outputs the survival prognostic risk index (such as OS / PFS score) for the nasopharyngeal carcinoma patient to the terminal.

[0049] The specific training process and principles of the pre-trained LASSO-Cox survival analysis model are as follows: Step 1: Component Data Transformation (Breaking Collinearity). Since the sum of the spatial proportions of the 53 MP-HPCs is 1 (i.e.... This data belongs to component data and exhibits strict negative correlation and collinearity. The model first performs a central log-ratio (CLR) transformation on it, mapping the constrained proportion data to an unconstrained real number space, resulting in a new 53-dimensional feature vector. .

[0050] Step Two: LASSO Feature Selection and Dimensionality Reduction. During model training, the survival time (e.g., OS or PFS) and survival status of patients in the training cohort are used as target variables. The aforementioned 53-dimensional CLR features are used as input, and a Cox proportional hazards regression model with L1 regularization penalty (i.e., LASSO-Cox) is used for training. The characteristic of the LASSO algorithm is that it compresses the regression coefficients of MP-HPCs features that contribute little to prognostic prediction or are redundant to absolute zero, thereby achieving feature selection and preventing overfitting.

[0051] Step 3: Determine weights and construct the risk score. After training, the model will retain K features (K≤53) with non-zero coefficients. Each of these K features has a non-zero Cox regression coefficient β (i.e., its own independent weight, distinguishing between positive and negative; positive weights represent risk factors, and negative weights represent protective factors). The final prognostic risk assessment model output formula is:

[0052] (in, Let j be the CLR transformation value of the j-th selected MP-HPCs. (This corresponds to the model weights). Based on the calculated Risk Score, the system stratifies patients' risk levels (high / low) end-to-end.

[0053] In some embodiments, the preset parameters or thresholds of the above design are obtained by training with training set samples. They can be specific thresholds or ranges. The specific form is not specifically limited in this embodiment.

[0054] This application also discloses a computer device, which may include: one or more processors and one or more memories; wherein the memories store computer-readable code that, when run by the one or more processors, can perform the methods described above.

[0055] The processor in this embodiment can be an integrated circuit chip with signal processing capabilities. The processor can be a general-purpose processor, a digital signal processor (DSP), an application-specific integrated circuit (ASIC), an off-the-shelf programmable gate array (FPGA), or other programmable logic devices, discrete gate or transistor logic devices, or discrete hardware components. It can implement or execute the methods, operations, and logic block diagrams disclosed in this embodiment. The general-purpose processor can be a microprocessor or any conventional processor, and can be based on an x86 or ARM architecture.

[0056] In general, the various exemplary embodiments of this disclosure can be implemented in hardware or dedicated circuitry, software, firmware, logic, or any combination thereof. Some aspects can be implemented in hardware, while others can be implemented in firmware or software that can be executed by a controller, microprocessor, or other computing device. When aspects of embodiments of this disclosure are illustrated or described as block diagrams, flowcharts, or using some other graphical representation, it will be understood that the blocks, apparatuses, systems, techniques, or methods described herein can be implemented as non-limiting examples in hardware, software, firmware, dedicated circuitry or logic, general-purpose hardware or controllers or other computing devices, or some combination thereof.

[0057] For example, the methods or apparatus according to embodiments of this disclosure can also be implemented using the architecture of a computing device. A computing device may include a bus, one or more CPUs, read-only memory (ROM), random access memory (RAM), a communication port connected to a network, input / output components, a hard disk, etc. Storage devices in the computing device, such as ROM or hard disk, may store various data or files used for processing and / or communication of the methods provided in this disclosure, as well as program instructions executed by the CPU. The computing device may also include a user interface. Of course, the architecture described above is merely exemplary, and one or more components of the computing device described above may be omitted as needed when implementing different devices.

[0058] This invention also provides a computer-readable storage medium storing computer-readable instructions. When the computer-readable instructions are executed by a processor, the methods disclosed in this embodiment can be performed. The computer-readable storage medium in this embodiment can be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. Non-volatile memory can be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. Volatile memory can be random access memory (RAM), which is used as an external cache. By way of example, but not limitation, many forms of RAM are available, such as static random access memory (SRAM), dynamic random access memory (DRAM), synchronous dynamic random access memory (SDRAM), double data rate synchronous dynamic random access memory (DDRSDRAM), enhanced synchronous dynamic random access memory (ESDRAM), synchronous interconnected dynamic random access memory (SLDRAM), and direct memory bus random access memory (DRRAM). It should be noted that the memory of the methods described herein is intended to include, but is not limited to, these and any other suitable types of memory.

[0059] This disclosure also provides a computer program product or system, including a computer program that, when executed by a processor, implements the steps of the above-described method.

[0060] like Figure 2 As shown, this embodiment also provides a dual-track parallel purification system for high-dimensional spatial transcriptome data, specifically including: The virtual whole transcriptome expression matrix acquisition module 201 is used or configured to acquire the virtual whole transcriptome expression matrix of image blocks in the histopathological whole slice image of the target disease. The virtual whole transcriptome expression matrix is ​​predicted and output by the generative spatial transcriptome prediction model of the image blocks. The target gene extraction module 202 is used or configured to screen a subset of target gene expression from the virtual whole transcriptome expression matrix based on a preset set of immune microenvironment-related genes. The dual-track parallel dimensionality reduction processing module 203 is used or configured to perform dual-track parallel dimensionality reduction processing on the target gene expression subset: In the first processing track, a core gene subset that meets the preset immune relevance is extracted from the target gene expression subset. The genes in the core gene subset are divided into N functional modules according to their biological functions. The statistical characteristics of gene expression in each functional module are calculated to generate a modular feature matrix. In the second processing track, nonnegative matrix factorization is performed on the expression levels of the remaining genes in the target gene expression subset to extract the basis matrix as the potential feature matrix of the microenvironment. The matrix merging module 204 is used or configured to merge the modular feature matrix and the microenvironment potential feature matrix to obtain a denoised pure virtual molecular feature matrix.

[0061] In some embodiments, the system further includes: A macroscopic morphological feature matrix acquisition module is used or configured to acquire the macroscopic morphological feature matrix of image blocks in the whole histopathological slice image, wherein the macroscopic morphological feature matrix is ​​obtained by a visual feature extraction model extracting features from the same image block; A multimodal fusion feature matrix generation module is used or configured to perform image patch-level vector concatenation of the pure virtual molecular feature matrix and the macroscopic morphological feature matrix obtained based on the same image patch to generate a multimodal fusion feature matrix; A multimodal morphology-molecular phenotype cluster generation module is used or configured to construct a nearest neighbor graph network based on the multimodal fusion feature matrix and perform clustering through a community detection algorithm to generate multimodal morphology-molecular phenotype clusters.

[0062] In some embodiments, the system further includes: The component vector calculation module is used or configured to calculate the proportion of image blocks of each category of the multimodal morphology-molecular phenotype cluster in a single histopathological whole slice image to obtain the component vector; A component vector transformation module is used or configured to perform a central logarithmic ratio transformation on the component vector to obtain a transformed feature vector; The prognostic risk score output module is used or configured to input the transformed feature vector into a pre-trained Cox survival analysis model to obtain a prognostic risk score. The Cox survival analysis model incorporates a regularization penalty mechanism during training. Using the survival time and survival status of patients in the training cohort as target variables, feature selection is performed on the transformed feature vector. The regression coefficients of features that contribute less than a preset threshold to prognostic prediction are compressed to zero, and only features with non-zero regression coefficients and their corresponding weights are retained. The prognostic risk score is obtained by linearly weighting the transformed values ​​of the retained features and their corresponding weights.

[0063] In this embodiment, nasopharyngeal carcinoma is used as the target disease to exemplify the description of a dual-track parallel purification system for high-dimensional spatial transcriptome data. However, this description is only one use case and is not intended to limit the scope of protection of this data processing method. Specifically, as shown in the example... Figure 3 The diagram illustrates the overall data flow and execution sequence of the feature extraction and splicing process in the method of this invention. First, the server acquires WSI slices from nasopharyngeal carcinoma patients, performs tissue segmentation and block preprocessing, and outputs standardized image blocks. Second, the data is split into a dual-branch network; the morphological branch outputs 128-dimensional macroscopic features, and the molecular branch outputs the predicted whole transcriptome expression matrix. Subsequently, the noise reduction and purification module performs mean normalization and NMF dimensionality reduction on the molecular data, outputting 60-dimensional purified features, which are then vector-spliced ​​with the morphological features. Finally, the spliced ​​features are combined into a 188-dimensional feature vector. A specific example is described below: Data acquisition and preprocessing module In this stage, the server mainly completes the physical segmentation and color normalization of the input data, providing a standard input source for subsequent deep learning networks.

[0064] Step 1: Obtain the whole-slide image (WSI) file of nasopharyngeal carcinoma pathology generated by the whole-field digital slide scanner.

[0065] Step 2: The server executes the tissue segmentation and dicing algorithm. At a magnification of 10X or 20X, the effective tissue region of WSI is divided into 224 x 224 pixel non-overlapping image blocks.

[0066] Step 3: Perform color standardization. The server uses the Reinhard algorithm to perform color space conversion on the extracted image patches, eliminating H&E staining color deviations caused by different institutions and scanners, and outputting a standardized set of image patches.

[0067] Dual-branch feature extraction module This module contains two parallel neural network processing streams. The server extracts two types of high-dimensional information—macroscopic tissue morphology and microscopic gene expression—simultaneously from the same standardized image patch.

[0068] Step 1 (Historical Morphology Branch Processing): The server calls a self-supervised visual foundation model (UNI Encoder based on the Vision Transformer architecture) pre-trained on massive amounts of pathological images. A single image patch is input into the model to extract deep network features, and the output dimension is... The initial morphological feature vector. To reduce computational complexity and remove redundant information, the server performs L2 regularization on this 1024-dimensional vector. After standardization (Z-scoreScaler), principal component analysis (PCA) is called to reduce the dimensionality of the feature matrix to [value missing]. Dimension, yielding a refined macroscopic morphological feature matrix. , where N is the total number of image blocks.

[0069] Step 2 (Gene Expression Branching): The server inputs the same image patch into the Generative Spatial Transcriptome Model (STPath). The STPath network learns the mapping relationship between tissue morphology and the spatial transcriptome, predicts the virtual gene expression levels of approximately 38,984 genes within the image patch region, and outputs a virtual whole transcriptome expression matrix. .

[0070] Virtual molecular feature denoising and purification module, such as Figure 4 The algorithm for virtual molecular feature denoising and multimodal fusion is illustrated. The server inputs the original molecular expression matrix of more than 30,000 dimensions from the same image patch into the filtering module, retaining 862 feature genes. Among them, 100 core genes are mapped to 16 prior biological functional modules, and after Z-score normalization and mean calculation, they are extracted into 32-dimensional feature vectors; the remaining 762 related genes are input into the NMF nonnegative matrix factorizer and extracted into 28-dimensional latent feature vectors.

[0071] Because the 38984-dimensional gene expression matrix output by STPath suffers from an extremely high curse of dimensionality, and a large number of genes have no biological association with immunotherapy response, the server performs dimensionality reduction and purification of the data stream using the following secondary algorithm: Step 1 (Prior Knowledge Screening): The server reads the preset gene dictionary of nasopharyngeal carcinoma immune microenvironment features, performs feature index filtering on the Fst matrix, and retains only 862 gene features that are highly related to nasopharyngeal carcinoma immunotherapy.

[0072] Step 2 (Dual-track dimensionality reduction algorithm): Split the 862 genes into two data streams for asynchronous processing.

[0073] Sub-path A (Functional module mean-based calculation): The server extracts 100 core genes and maps them to 16 functional modules (such as T cell infiltration, stem T cells, TLS chemotaxis and structural modules, etc.) according to biological prior logic. For each module... Gene expression levels within the cell were first normalized using Z-scores. (in Gene expression level, and (This represents the mean and standard deviation of gene j). Then, the mean and coefficient of variation of all genes within this module are calculated. Finally, each module is condensed into two statistical features, and the combined output is... Modular feature matrix of dimension The physical significance of this calculation formula lies in smoothing out the prediction noise of a single gene and improving the noise resistance of key immune pathways.

[0074] Sub-path B (non-negative matrix factorization mining): The server processes the remaining 762 gene expression matrices. Perform an unsupervised nonnegative matrix factorization (NMF) algorithm. The algorithm sets the rank of the factorization. By iteratively optimizing the objective function (W ≥ 0, H ≥ 0), find the basis matrix W and coefficient matrix H, and automatically extract potential and unknown microenvironment gene co-expression patterns. The server extracts matrix W as... 3D microenvironment latent feature matrix Ultimately, the high-dimensional noise of the molecular branch was significantly compressed, resulting in a pure virtual molecular feature output with a dimension of 60 (32+28).

[0075] Multimodal feature fusion and graph clustering module, such as Figure 5 As shown, the differences in the output maps of unimodal and multimodal clustering networks are compared. Before the injection of molecular features (left side), the node network only uses K-nearest neighbor mapping based on 128-dimensional morphological features, and the Leiden algorithm divides the tissue morphology into 35 HPCs (tissue morphology phenotypic clusters). After the introduction of microscopic molecular features (right side), the network topology changes. Nodes that were originally highly morphologically similar (such as HPC 14 and 18) generate internal repulsive forces, are successfully torn apart, and reconstructed into 53 more fine-grained multimodal morphology-molecular phenotypic clusters (MP-HPCs), effectively breaking through the visual blind spot of homomorphic heterogeneity in traditional H&E slices.

[0076] In this stage, the server performs low-level fusion of data streams from different modalities and completes the construction of phenotypic maps.

[0077] Step 1 (Vector Concatenation): The server concatenates the morphological matrix at the image patch level. (128-dimensional) and the denoised molecular feature matrix (32-dimensional) (28-dimensional) Perform splicing operation to generate Multimodal fusion feature matrix .

[0078] Step 2 (Graph Network Construction and Leiden Clustering): The server calculates the Euclidean distance between each pair of image patches on the 188-dimensional multimodal matrix, constructing a K-Nearest Neighbors (KNN) graph network. Subsequently, the Leiden community detection algorithm is called to perform unsupervised clustering optimization on the network. The internal implementation principle of the Leiden algorithm is to maximize the modularity H of the graph, and its calculation formula is:

[0079] The physical meanings of the formula parameters are as follows: k represents the actual number of edges assigned to cluster (community) c. C Let represent the sum of the degrees of all nodes in cluster c (i.e., the expected number of edges), and m be the total number of edges in the network. γ is the resolution parameter used to control the granularity of clustering. Algorithm execution sequence: First, each node is treated as an independent community. Then, the modularity gain when moving a node to an adjacent community is iteratively evaluated. If the gain is positive, the node is moved until the network reaches a local optimum.

[0080] Step 3 (Clustering Reconstruction Effect): When using only 128-dimensional morphological features, Leiden clustering outputs 35 basic tissue morphological phenotypic clusters (HPCs). This scheme, by injecting 60-dimensional molecular signals, alters the topological connectivity of the KNN graph network. Originally morphologically similar HPC nodes are automatically split and reconstructed under the repulsive force of molecular features, ultimately outputting 53 more detailed morphology-molecular multimodal phenotypic clusters (MP-HPCs).

[0081] Prognostic risk assessment and output module For end-to-end clinical applications, the server performs survival prediction assessments based on the calculated MP-HPCs atlas, such as... Figure 6 The diagram illustrates the end-to-end prognostic risk assessment system architecture. The terminal extracts the area proportion components of 53 MP-HPCs from a single patient's WSI, constructing a component data vector. The server performs a central log-ratio (CLR) transformation on this vector to eliminate feature collinearity, and then inputs it into a pre-trained LASSO-Cox risk assessment model. The model calculates the SHAP-weighted sum of the contributions of each MP-HPC, ultimately outputting the patient's prognostic risk index to guide clinical decisions regarding immunotherapy regimens.

[0082] Step 1: Count the number of image patches belonging to each MP-HPC in a single patient's WSI, divide by the total number of image patches in that WSI, and obtain the MP-HPC spatial proportion component vector with dimension 53. ,in .

[0083] Step 2: To eliminate collinearity of component data and satisfy the independence assumption of Cox regression, the server executes the Center Log-Ratio (CLR) transformation algorithm. The formula principle is as follows:

[0084] in, The geometric mean of the component vectors: .

[0085] Step 3: Input the transformed features into the pre-trained LASSO-Cox survival analysis model. The model calculates the sum of risks based on the predefined regression coefficients of each MP-HPC and outputs the survival prognostic risk index (such as OS / PFS score) for the nasopharyngeal carcinoma patient to the terminal.

[0086] The exemplary embodiments of this disclosure described in detail above are merely illustrative and not restrictive. Those skilled in the art will understand that various modifications and combinations of these embodiments or their features can be made without departing from the principles and spirit of this disclosure, and such modifications should fall within the scope of this disclosure.

Claims

1. A dual-track parallel purification method for high-dimensional spatial transcriptome data, characterized in that, include: A virtual whole transcriptome expression matrix is ​​obtained from the image patch in the whole tissue pathology image of the target disease. The virtual whole transcriptome expression matrix is ​​predicted and output by the generative spatial transcriptome prediction model for the image patch. Based on a preset set of immune microenvironment-related genes, a subset of target gene expressions is selected from the virtual whole transcriptome expression matrix; The target gene expression subset is subjected to dual-track parallel dimensionality reduction processing: in the first processing track, a core gene subset that meets the preset immune correlation is extracted from the target gene expression subset, and the genes in the core gene subset are divided into N functional modules according to their biological functions. The statistical characteristics of gene expression in each functional module are calculated to generate a modular feature matrix. In the second processing track, nonnegative matrix factorization is performed on the expression levels of the remaining genes in the target gene expression subset to extract the basis matrix as the potential feature matrix of the microenvironment. The modular feature matrix and the microenvironment latent feature matrix are combined to obtain the denoised pure virtual molecular feature matrix.

2. The dual-track parallel purification method for high-dimensional spatial transcriptome data according to claim 1, characterized in that, The preset immune microenvironment-related gene set includes genes related to the immunotherapy response to the target disease; Optionally, the core gene subset includes M key biomarker genes that meet the preset immune-related criteria, where M is a natural number greater than 5.

3. The dual-track parallel purification method for high-dimensional spatial transcriptome data according to claim 1, characterized in that, The functional modules are divided according to immune cell subsets or immune pathway functions, where N is a natural number greater than or equal to 2, and includes at least two of the following modules: T cell infiltration module, T cell exhaustion module, and macrophage planning module.

4. The dual-track parallel purification method for high-dimensional spatial transcriptome data according to claim 1, characterized in that, The statistical features include the mean and / or coefficient of variation calculated after standardizing the expression values ​​of each gene within the functional module.

5. The dual-track parallel purification method for high-dimensional spatial transcriptome data according to claim 1, characterized in that, The decomposition rank of the non-negative matrix factorization in the second processing track is a settable parameter, and a preset number of potential gene co-expression patterns are automatically extracted by performing non-negative matrix factorization.

6. The dual-track parallel purification method for high-dimensional spatial transcriptome data according to claim 1, characterized in that, The method further includes: obtaining the macroscopic morphological feature matrix of the image patch in the whole tissue pathology slice image of claim 1, wherein the macroscopic morphological feature matrix is ​​obtained by extracting features from the same image patch using a visual feature extraction model; and performing image patch-level vector concatenation of the pure virtual molecular feature matrix and the macroscopic morphological feature matrix obtained based on the same image patch to generate a multimodal fusion feature matrix. A nearest neighbor graph network is constructed based on the multimodal fusion feature matrix, and clustering is performed using a community detection algorithm to generate multimodal morphological-molecular phenotypic clusters.

7. The dual-track parallel purification method for high-dimensional spatial transcriptome data according to claim 1, characterized in that, The method also includes: The proportion of image patches for each category of multimodal morphology-molecular phenotype clusters in a single histopathological whole-section image is calculated to obtain the component vector; Perform a central logarithmic ratio transformation on the component vector to obtain the transformed feature vector; The transformed feature vector is input into a pre-trained Cox survival analysis model to obtain a prognostic risk score; The Cox survival analysis model incorporates a regularization penalty mechanism during training. Using the survival time and survival status of patients in the training cohort as target variables, feature selection is performed on the transformed feature vector. The regression coefficients of features that contribute less than a preset threshold to prognostic prediction are compressed to zero, and only features with non-zero regression coefficients and their corresponding weights are retained. The prognostic risk score is obtained by linearly weighting the transformed values ​​of the retained features and their corresponding weights.

8. A computer device, characterized in that, The device includes: a memory and a processor; the memory is used to store a computer program; the processor executes the computer program to implement the steps of the method according to any one of claims 1-7.

9. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the steps of the method as described in any one of claims 1-7.

10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the method described in any one of claims 1-7.

Images