Metastasis prediction method, device, medium and equipment for colorectal cancer based on spatial omics

Abstract

The application discloses a kind of based on space omics's intestinal cancer metastasis prediction method, device, medium and equipment, method includes: acquisition original multi-omics data, modal alignment and quality control processing are handled, obtain including second spatial transcriptome data, second single-cell RNA sequencing data and second pathology image data pre-processing multi-omics data;Second single-cell RNA sequencing data is based on second spatial transcriptome data and is carried out cross-modal semantic embedding, generates space enhanced expression profile;Second spatial transcriptome data and second pathology image data are carried out multi-scale graph construction, and spatial heterogeneity feature is extracted;Space enhanced expression profile and spatial heterogeneity feature are input into pre-trained metastasis risk prediction model, and output liver metastasis probability space heat map and key driving feature list;Finally generate the clinical prediction report containing high-risk area positioning.The application significantly improves the sensitivity of early metastasis detection by dynamically optimizing spatial resolution and multi-scale feature collaborative modeling.

Description

Technical Field

[0001] This invention relates to the field of medical technology, specifically to a method, apparatus, medium, and device for predicting colorectal cancer metastasis based on spatial omics. Background Technology

[0002] Early prediction of colorectal cancer metastasis is crucial for clinical treatment. Traditional pathological examination relies on morphological observation, which can identify some high-risk features but struggles to quantify the spatial heterogeneity of the tumor microenvironment. Molecular detection techniques, such as gene expression profiling, can provide molecular characteristics of tumors but lose crucial spatial information, failing to reflect regional differences within the tumor. Single-cell sequencing improves resolution but still cannot preserve information on the spatial distribution of cells in situ within the tissue.

[0003] In recent years, spatial omics technology has made it possible to simultaneously acquire gene expression and its spatial distribution. However, current spatial transcriptomics detection has limited resolution, and individual data points may contain mixed cellular signals, affecting the identification of micrometastases. Furthermore, how to effectively integrate multi-scale data such as molecular expression, cell interactions, and tissue structure, and establish high-precision predictive models, remains a pressing technical challenge. Existing prediction methods often rely on single data points or simple feature aggregation, failing to fully utilize the complementarity of multi-source data, resulting in insufficient accuracy in predicting early metastasis risk. Summary of the Invention

[0004] In view of the above problems, the present invention provides a method, device, medium and equipment for predicting colorectal cancer metastasis based on spatial omics. By dynamically optimizing spatial resolution and co-modeling multi-scale features, it achieves high sensitivity in detecting metastasis risk and solves the problem of high false negative rate in early metastasis.

[0005] To achieve the above objectives, in a first aspect, this application provides a method for predicting colorectal cancer metastasis based on spatial omics, comprising:

[0006] Raw multi-omics data were collected from original colorectal cancer tissue samples. The raw multi-omics data included first spatial transcriptome data, first single-cell RNA sequencing data, and first pathological image data.

[0007] Modality alignment and quality control were performed on the raw multi-omics data to obtain preprocessed multi-omics data, which included second spatial transcriptome data, second single-cell RNA sequencing data, and second pathological image data.

[0008] Based on second single-cell RNA sequencing data, cross-modal semantic embedding was performed on second spatial transcriptome data to generate spatially enhanced expression profiles. Cross-modal semantic embedding included single-cell-spatial feature alignment and probabilistic sparse signal completion based on contrastive learning. The spatially enhanced expression profiles included cell type deconvolution results and confidence scores of metastasis-related genes.

[0009] Multiscale graphs were constructed from the second spatial transcriptome data and the second pathological image data to extract spatial heterogeneity features. The multiscale graph construction included microscopic cell interaction graphs, mesoscopic functional region graphs, and hierarchical modeling of macroscopic tissue structures. Spatial heterogeneity features included tumor-immune adjacency strength, spatial gradient of the vascular invasion front, and topological relationships of high-risk subregions.

[0010] The spatially enhanced expression profile and spatial heterogeneity features are input into the pre-trained metastasis risk prediction model, and the output is a spatial heatmap containing the probability of liver metastasis and a list of key driving features.

[0011] Based on spatial heatmaps and a list of key driving features, a clinical prediction report on colorectal cancer metastasis risk is generated. The report includes high-risk area location, metastasis probability score, and treatment response prediction.

[0012] Furthermore, modality alignment and quality control processing are performed on the original multi-omics data to obtain preprocessed multi-omics data, including:

[0013] Spatial sites that meet the preset quality thresholds are selected from the first spatial transcriptome data as the second spatial transcriptome data. The preset quality thresholds include the gene detection threshold and the upper limit of the proportion of mitochondrial genes.

[0014] Cell sequencing data that match the anatomical regions of the first single-cell RNA sequencing data with the second spatial transcriptome data are extracted as the second single-cell RNA sequencing data. Anatomical region matching is achieved through tissue slice coordinate mapping.

[0015] The image region corresponding to the acquisition location of the second spatial transcriptome data is selected from the first pathological image data as the second pathological image data, and the spatial deviation of the acquisition location during matching does not exceed the preset alignment tolerance.

[0016] Establish spatial correlations among second spatial transcriptome data, second single-cell RNA sequencing data, and second pathological image data, so that the three have a unified coordinate reference system;

[0017] Batch effect correction was performed on the second-space transcriptome data to eliminate technical variations between different samples;

[0018] Low-quality cells were filtered from the second single-cell RNA sequencing data to retain transcriptome data that met cell integrity standards.

[0019] Generate preprocessed multi-omics data, and store multiple preprocessed multi-omics data in a quality control database.

[0020] Furthermore, cross-modal semantic embedding of the second spatial transcriptome data is performed based on the second single-cell RNA sequencing data to generate a spatially enhanced expression profile, including:

[0021] Cell type feature matrix was extracted from the second single-cell RNA sequencing data. The cell type feature matrix contains the expression profile of marker genes for each cell type.

[0022] Each spatial site in the second spatial transcriptome data is compared and aligned with the cell type feature matrix, and the similarity score between the spatial site and each cell type is calculated.

[0023] Based on similarity scores, probabilistic cell type deconvolution is performed on spatial sites to generate a cell type ratio matrix. The deconvolution is jointly optimized using nonnegative matrix factorization and Markov random fields.

[0024] Sparse signal completion was performed on low-expression genes in the second space transcriptome data. The completion signal was derived from the prior expression distribution of the same type of cells in single-cell data.

[0025] By fusing the cell type ratio matrix with the completed gene expression data, a spatially enhanced expression profile is generated, which includes cell composition information and gene expression confidence scores for each spatial site.

[0026] Furthermore, based on similarity scores, probabilistic cell type deconvolution is performed on spatial sites to generate a cell type ratio matrix. The deconvolution is jointly optimized using nonnegative matrix factorization and Markov random fields, including:

[0027] The similarity score is used as the initial weight to construct a non-negative matrix factorization objective function of spatial site-cell type. Based on the spatial coordinate information of the second spatial transcriptome data, a neighborhood graph of the Markov random field is established. The neighborhood graph contains the topological connection relationship between spatial sites.

[0028] A spatial smoothing constraint term is introduced into the nonnegative matrix factorization objective function. The spatial smoothing constraint term adjusts the difference in cell type ratio between adjacent sites through a Markov random field.

[0029] The cell type proportion matrix is ​​updated iteratively using an alternating optimization algorithm. Each iteration includes:

[0030] With fixed space constraints, the cell type ratio of the non-negative matrix factorization is optimized by multiplication update rules; and with fixed cell type ratios, spatial consistency is optimized by graph Laplacian regularization.

[0031] The optimization terminates when the change in the Frobenius norm of the cell type ratio matrix in adjacent iterations is less than the preset convergence threshold.

[0032] The optimized cell type ratio matrix is ​​normalized to ensure that the sum of the cell type ratios at each spatial location is 1;

[0033] Output a cell type ratio matrix. The rows of the cell type ratio matrix correspond to spatial locations, the columns correspond to cell types, and the element values ​​are probabilistic ratio weights.

[0034] Furthermore, multi-scale maps were constructed from the second spatial transcriptome data and the second pathological image data to extract spatial heterogeneity features, including:

[0035] Based on single-cell resolution expression profiles from second-space transcriptome data, a micro-cell interaction network diagram was constructed. This micro-cell interaction network diagram was generated through ligand-receptor pair co-expression analysis and cell adjacency calculation.

[0036] Tissue region segmentation was performed on the second pathological image data to identify the tumor core area, invasion front area and mesenchymal area. Combined with the expression characteristics of the second spatial transcriptome data of the corresponding regions, a mesoscopic functional region map was constructed.

[0037] By integrating the tissue structure features of second pathological image data at the whole-slice level with the global expression patterns of spatial transcriptome data, a macroscopic tissue metastasis trend prediction map was established.

[0038] Extract tumor-immune cell interaction strength features from microscopic cell interaction network diagrams, including the spatial colocalization frequency of immune checkpoint molecular pairs.

[0039] The spatial gradient characteristics of the vascular invasion front were calculated from the mesoscopic functional region atlas, including the attenuation coefficient of endothelial marker expression with distance.

[0040] Topological features of high-risk sub-regions were extracted from the macro-organization migration trend prediction map, including the spatial autocorrelation index of the microenvironment before migration.

[0041] The characteristics of tumor-immune cell interaction intensity, spatial gradient of vascular invasion front, and topological characteristics of high-risk subregions are integrated into spatial heterogeneity features.

[0042] Furthermore, by integrating the structural features of second pathological image data at the whole-slice level with the global expression patterns of spatial transcriptome data, a macroscopic tissue metastasis trend prediction map is established, including:

[0043] Full-slice scanning was performed on the second pathological image data to extract tissue structural features, including the orientation of tumor glands, the density of stromal fibrosis distribution, and the spatial proportion of necrotic areas.

[0044] Spatial autocorrelation analysis was performed on the global expression patterns of the second spatial transcriptome data to calculate the spatial clustering index of expression hotspots of each gene. The global expression patterns included the spatial distribution of expression levels of epithelial-mesenchymal transition markers, angiogenesis factors, and immunosuppression-related molecules.

[0045] A multimodal fusion model is constructed based on graph convolutional neural networks. It aligns tissue structure features with global expression patterns. The alignment process uses a cross-modal attention mechanism to allocate weights between pathological image features and transcriptome expression features, and outputs a fused multimodal feature matrix.

[0046] A macroscopic tissue metastasis trend prediction map is established in the multimodal feature matrix. The macroscopic tissue metastasis trend prediction map simulates the potential metastasis path of tumor cells along tissue structure features and expression gradient features through a random walk algorithm.

[0047] Topology optimization is performed on the macro-organization transfer trend prediction map to identify high-risk transfer sub-regions. Topology optimization includes calculating the spatial density of transfer path intersections and the consistency coefficient of transfer direction.

[0048] The final output is a macro-organizational migration trend prediction map, which includes a migration probability heatmap, high-risk sub-region boundary markings, and a vector field of major migration paths.

[0049] Furthermore, the spatially enhanced expression profile and spatial heterogeneity features are input into the pre-trained metastasis risk prediction model, which outputs a spatial heatmap containing liver metastasis probability and a list of key driving features, including:

[0050] Feature normalization was performed on the spatially enhanced expression profile to make the expression levels of different genes comparable, including Z-score normalization and batch effect correction.

[0051] Spatial heterogeneity features are dimensionally compressed, and principal component analysis is used to retain principal component features whose contribution rate exceeds a preset threshold.

[0052] The standardized spatial augmented expression spectrum and the compressed spatial heterogeneous features are concatenated to generate a fused feature matrix;

[0053] The fused feature matrix is ​​input into a pre-trained graph neural network model, which aggregates multi-scale spatial information based on an attention mechanism.

[0054] The probability of liver metastasis at each spatial site is calculated using a graph neural network model, and a spatial heatmap is generated based on the probability value. The spatial heatmap maintains a spatial correspondence with the original tissue slices.

[0055] The gradient backpropagation method is used to extract key driving features from the graph neural network model, including gene expression features and spatial topology features that contribute the most to the prediction results;

[0056] The key driving features are sorted according to their contribution, and a list of key driving features including feature name, contribution score and biological explanation is generated.

[0057] The spatial heatmap is associated with and stored in a list of key driving features, and a mapping relationship is established with clinicopathological parameters.

[0058] In a second aspect, the present invention also provides a spatial omics-based colorectal cancer metastasis prediction device, applicable to the method described in the first aspect. The device includes: a data acquisition module, a data processing module, a logic operation module, and a report generation module. The data acquisition module is used to acquire raw multi-omics data from original colorectal cancer tissue samples. The raw multi-omics data includes first spatial transcriptome data, first single-cell RNA sequencing data, and first pathological image data. The data processing module is used to perform modality alignment and quality control processing on the raw multi-omics data to obtain preprocessed multi-omics data. The preprocessed multi-omics data includes second spatial transcriptome data, second single-cell RNA sequencing data, and second pathological image data. The logic operation module is used to perform cross-modal semantic embedding on the second spatial transcriptome data based on the second single-cell RNA sequencing data to generate a spatially enhanced expression profile. The cross-modal semantic embedding includes... The model employs single-cell spatial feature alignment and probabilistic sparse signal completion, along with spatially enhanced expression profiling that includes cell type deconvolution results and confidence scores for metastasis-related genes. Multi-scale graphs are constructed from second-dimensional transcriptome data and second-dimensional pathological image data to extract spatial heterogeneity features. These multi-scale graphs include microscopic cell interaction maps, mesoscopic functional region maps, and hierarchical modeling of macroscopic tissue structures. Spatial heterogeneity features include tumor-immune adjacency strength, spatial gradients at the vascular invasion front, and topological relationships in high-risk subregions. The spatially enhanced expression profiling and spatial heterogeneity features are input into a pre-trained metastasis risk prediction model, outputting a spatial heatmap containing liver metastasis probability and a list of key driving features. A report generation module generates a clinical prediction report of colorectal cancer metastasis risk based on the spatial heatmap and the list of key driving features. The report includes high-risk region location, metastasis probability score, and treatment response prediction.

[0059] In a third aspect, the present invention also provides a computer-readable storage medium having stored thereon computer program instructions that, when executed by a processor, implement the method described in the first aspect.

[0060] In a fourth aspect, the present invention also provides an electronic device including a memory and a processor, the memory being used to store one or more computer program instructions, wherein the one or more computer program instructions are executed by the processor to implement the method described in the first aspect.

[0061] Unlike existing technologies, the above-mentioned technical solution provides a method, device, medium, and equipment for predicting colorectal cancer metastasis based on spatial omics. The method includes: collecting raw multi-omics data, performing modality alignment and quality control processing to obtain preprocessed multi-omics data including second spatial transcriptome data, second single-cell RNA sequencing data, and second pathological image data; performing cross-modal semantic embedding on the second spatial transcriptome data based on the second single-cell RNA sequencing data to generate a spatially enhanced expression profile; constructing multi-scale maps on the second spatial transcriptome data and the second pathological image data to extract spatial heterogeneity features; inputting the spatially enhanced expression profile and spatial heterogeneity features into a pre-trained metastasis risk prediction model to output a spatial heatmap of liver metastasis probability and a list of key driving features; and finally generating a clinical prediction report containing the location of high-risk regions. This invention effectively identifies micro-metastasis signals by dynamically optimizing spatial resolution; and accurately quantifies the spatial heterogeneity features of the tumor microenvironment through multi-scale feature collaborative modeling, which can significantly improve the sensitivity and spatial positioning accuracy of colorectal cancer metastasis prediction.

[0062] The above description of the invention is merely an overview of the technical solution of this application. In order to enable those skilled in the art to better understand the technical solution of this application and to implement it based on the description and drawings, and to make the above-mentioned objectives and other objectives, features and advantages of this application easier to understand, the following description is provided in conjunction with the specific embodiments and drawings of this application. Attached Figure Description

[0063] The accompanying drawings are only used to illustrate the principles, implementation methods, applications, features, and effects of specific embodiments of the present invention and other related contents, and should not be considered as limitations on this application.

[0064] In the accompanying drawings of the instruction manual:

[0065] Figure 1 This is a flowchart illustrating steps S101 to S106 of the prediction method described in a specific implementation.

[0066] Figure 2 This is a flowchart illustrating steps S201 to S207 of the prediction method described in a specific implementation.

[0067] Figure 3 This is a flowchart illustrating steps S301 to S305 of the prediction method described in a specific implementation.

[0068] Figure 4 This is a flowchart illustrating steps S401 to S403 of the prediction method described in a specific implementation.

[0069] Figure 5 This is a schematic diagram of the structure of the electronic device described in a specific embodiment.

[0070] The reference numerals used in the above figures are explained as follows:

[0071] 1. Electronic equipment;

[0072] 11. Memory;

[0073] 12. Processor. Detailed Implementation

[0074] To illustrate the possible application scenarios, technical principles, implementable specific solutions, and achievable objectives and effects of this application in detail, the following description, in conjunction with the listed specific embodiments and accompanying drawings, provides a detailed explanation. The embodiments described herein are merely illustrative of the technical solutions of this application and are therefore intended to limit the scope of protection of this application.

[0075] In this document, the term "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The term "embodiment" appearing in various places throughout the specification does not necessarily refer to the same embodiment, nor does it specifically limit its independence or connection with other embodiments. In principle, in this application, as long as there are no technical contradictions or conflicts, the technical features mentioned in each embodiment can be combined in any way to form corresponding implementable technical solutions.

[0076] Unless otherwise defined, the technical terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains; the use of related terms herein is merely for the purpose of describing particular embodiments and is not intended to limit this application.

[0077] In the description of this application, the term "and / or" is used to describe the logical relationship between objects, indicating that three relationships can exist. For example, A and / or B means: A exists, B exists, and A and B exist simultaneously. Additionally, the character " / " in this document generally indicates that the preceding and following objects have an "or" logical relationship.

[0078] In this application, terms such as “first” and “second” are used only to distinguish one entity or operation from another, and do not necessarily require or imply any actual quantity, hierarchy or order relationship between these entities or operations.

[0079] Without further limitations, the use of terms such as “comprising,” “including,” “having,” or other similar open-ended expressions in this application is intended to cover non-exclusive inclusion, which does not exclude the presence of additional elements in a process, method, or product that includes the stated elements, such that a process, method, or product that includes a list of elements may include not only those defined elements but also other elements not expressly listed, or elements inherent to such a process, method, or product.

[0080] As understood in the Examination Guidelines, in this application, expressions such as "greater than," "less than," and "exceeding" are understood to exclude the stated number; expressions such as "above," "below," and "within" are understood to include the stated number. Furthermore, in the description of the embodiments in this application, "multiple" means two or more (including two), and similar expressions related to "multiple" are also understood in this way, such as "multiple groups" and "multiple times," unless otherwise explicitly specified.

[0081] In the description of the embodiments of this application, the space-related expressions used, such as "center," "longitudinal," "lateral," "length," "width," "thickness," "upper," "lower," "front," "rear," "left," "right," "vertical," "horizontal," "vertical," "top," "bottom," "inner," "outer," "clockwise," "counterclockwise," "axial," "radial," and "circumferential," indicate the orientation or positional relationship based on the orientation or positional relationship shown in the specific embodiments or drawings. They are only for the purpose of describing the specific embodiments of this application or for the reader's understanding, and do not indicate or imply that the device or component referred to must have a specific position, a specific orientation, or be constructed or operated in a specific orientation. Therefore, they should not be construed as limitations on the embodiments of this application.

[0082] Please see Figure 1 In a first aspect, this embodiment provides a method for predicting colorectal cancer metastasis based on spatial omics, including:

[0083] S101. Collect raw multi-omics data from original colorectal cancer tissue samples. The raw multi-omics data includes first spatial transcriptome data, first single-cell RNA sequencing data, and first pathological image data.

[0084] S102. Modal alignment and quality control processing are performed on the raw multi-omics data to obtain preprocessed multi-omics data, which includes second spatial transcriptome data, second single-cell RNA sequencing data and second pathological image data.

[0085] S103. Based on the second single-cell RNA sequencing data, cross-modal semantic embedding is performed on the second spatial transcriptome data to generate a spatially enhanced expression profile. The cross-modal semantic embedding includes single-cell-spatial feature alignment and probabilistic sparse signal completion based on contrastive learning. The spatially enhanced expression profile includes cell type deconvolution results and confidence scores of transfer-related genes.

[0086] S104. Multi-scale graph construction was performed on the second spatial transcriptome data and the second pathological image data to extract spatial heterogeneity features. The multi-scale graph construction included microscopic cell interaction graphs, mesoscopic functional region graphs and hierarchical modeling of macroscopic tissue structures. Spatial heterogeneity features included tumor-immune adjacency strength, spatial gradient of the vascular invasion front and topological relationships of high-risk subregions.

[0087] S105. Input the spatially enhanced expression profile and spatial heterogeneity features into the pre-trained metastasis risk prediction model, and output a spatial heatmap containing the probability of liver metastasis and a list of key driving features.

[0088] S106. Based on spatial heatmaps and a list of key driving features, generate a clinical prediction report on colorectal cancer metastasis risk. The report includes high-risk area location, metastasis probability score, and treatment response prediction.

[0089] In step S101, the first spatial transcriptome data is a gene expression profile that preserves the spatial location of the tissue and can be obtained through a commercial spatial transcriptome sequencing platform; the first single-cell RNA sequencing data is the transcriptome data of a single tumor microenvironment cell, prepared using a clinical-grade single-cell isolation platform to ensure a sufficient number of epithelial cells and immune cells for subsequent analysis; the first pathological image data is a digitally scanned H&E-stained slide with a resolution sufficient for clear identification of cell nuclei for pathological diagnosis. It is important to note that the simultaneous acquisition of raw multi-omics data is achieved through a sample identification system to achieve spatiotemporal matching, and the pretreatment of tissue samples must be completed within a specified time after ex vivo to maintain RNA integrity.

[0090] In step S102, modal alignment maps data obtained from different detection technologies to a unified coordinate system, achieving precise registration between the spatial transcriptome chip and tissue sections through tissue structure markers in pathological images. During quality control processing, preferably, the second spatial transcriptome data is used to confirm RNA quality through internal reference gene detection, the second single-cell RNA sequencing data is used to filter low-quality cells and two-cell signals, and the second pathological image data is used to verify scan integrity and staining uniformity. Furthermore, the preprocessed data must meet the minimum quality control standards for downstream analysis; for example, a specified number of effective genes must be detected at each detection point in the spatial transcriptome.

[0091] In step S103, single-cell-spatial feature alignment maps gene expression patterns at single-cell resolution to spatial transcriptome sites using a similarity metric based on cell type features. Probabilistic sparse signal completion uses single-cell data to infer undetected gene expression in the spatial transcriptome, with its prior distribution set referencing known molecular features of colorectal cancer metastasis-related pathways. In the generated spatial enhanced expression profile, cell type deconvolution results show the proportional distribution of different cell types at each spatial site, and the confidence score for metastasis-related genes is the association strength between genes and metastatic phenotypes assessed through cross-validation.

[0092] In step S104, a microscopic cell interaction map depicts the spatial relationship between a single tumor cell and neighboring immune cells, achieved by calculating internuclear distance and contact area; a mesoscopic functional region map divides tumor subregions with similar gene expression patterns, which can be generated using a graph-based clustering algorithm; and a macroscopic tissue structure map analyzes the overall tumor growth pattern and invasion characteristics, aligning with the vascular invasion assessment criteria in pathological diagnosis. Among the extracted spatial heterogeneity features, tumor-immune adjacency strength is used to quantify the degree of immune cell encirclement of tumor cells, and the spatial gradient of the vascular invasion front measures the variation pattern of metastasis-related gene expression along blood vessels.

[0093] In step S105, the metastasis risk prediction model refers to a deep learning system trained on clinical data, whose input layer simultaneously receives gene expression features and spatial topological features. The output spatial heatmap visualizes the metastasis probability as a two-dimensional projection of the original tissue location, and its generation uses interpolation algorithms commonly used in medical image analysis; the key driving feature list consists of the molecular and spatial features that contribute the most to the prediction results, and is preferably extracted using a feature importance ranking algorithm and verified by pathology experts.

[0094] In step S106, high-risk area localization refers to mapping areas exceeding clinical thresholds in the spatial heatmap to pathological slide coordinates; the metastasis probability score is a composite risk assessment value that integrates molecular and spatial features, and its grading criteria are aligned with the metastasis risk stratification in clinical guidelines; treatment response prediction is evidence-based medicine that matches targeted treatment plans based on key driving features.

[0095] This embodiment leverages single-cell resolution data to enhance the analytical capabilities of spatial transcriptomics. It captures the heterogeneity of the tumor microenvironment through multi-scale spatial modeling guided by pathological images, ultimately fusing molecular expression patterns and spatial topological relationships for metastasis risk assessment. By integrating spatial transcriptomics data, single-cell RNA sequencing data, and pathological image data, it achieves multi-dimensional and accurate prediction of colorectal cancer metastasis risk. The spatially enhanced expression profiling addresses the insufficient resolution of traditional spatial transcriptomics data. Multi-scale map construction enables cross-level feature extraction from cell interactions to tissue structures. The clinical prediction report organically combines molecular features with pathological morphology, providing a comprehensive solution for colorectal cancer metastasis risk assessment that combines spatial resolution and clinical interpretability, thus helping to guide individualized treatment decisions.

[0096] Please see Figure 2 In some embodiments, modality alignment and quality control are performed on the raw multi-omics data to obtain preprocessed multi-omics data, including:

[0097] S201. Select spatial sites that meet the preset quality threshold from the first spatial transcriptome data as the second spatial transcriptome data. The preset quality threshold includes the gene detection threshold and the upper limit of the proportion of mitochondrial genes.

[0098] S202. Extract the cell sequencing data from the first single-cell RNA sequencing data that matches the anatomical region of the second spatial transcriptome data as the second single-cell RNA sequencing data. Anatomical region matching is achieved through tissue slice coordinate mapping.

[0099] S203. Select the image region corresponding to the acquisition location of the second spatial transcriptome data from the first pathological image data as the second pathological image data. The spatial deviation of the acquisition location during matching shall not exceed the preset alignment tolerance.

[0100] S204. Establish spatial correlations among second spatial transcriptome data, second single-cell RNA sequencing data, and second pathological image data, so that the three have a unified coordinate reference system;

[0101] S205. Perform batch effect correction on the second space transcriptome data to eliminate technical variations between different samples;

[0102] S206. Low-quality cells are filtered from the second single-cell RNA sequencing data to retain transcriptome data that meets the cell integrity standard.

[0103] S207. Generate preprocessed multi-omics data, and store multiple preprocessed multi-omics data in the quality control database.

[0104] In step S201, a preset quality threshold is used to assess the quality of spatial site data. The gene detection threshold is determined through the sequencing platform's technical white paper, representing the minimum number of genes that should be detected at each site. The upper limit for the proportion of mitochondrial genes is dynamically adjusted based on cell viability testing results to identify the degree of cell damage. This step retains biologically significant second-space transcriptome data through quality filtering while removing noise signals caused by technical factors.

[0105] In step S202, anatomical region matching spatially maps the single-cell sample origin location to spatial transcriptome chip coordinates. Preferably, coordinate mapping is achieved using an affine transformation algorithm in a digital pathology system. This process requires correction for deformation errors caused by tissue section processing. The extracted second single-cell RNA sequencing data ensures the inclusion of the main cell subpopulations of the target region, providing an accurate single-cell reference atlas for subsequent spatial analysis.

[0106] In step S203, the preset alignment tolerance refers to the maximum allowable positional deviation between the pathological image and the spatial transcriptome data. It is calculated and determined based on the tissue slice thickness and image resolution. The selected second pathological image data completely covers the detection area of ​​the spatial transcriptome. Preferably, sub-millimeter alignment accuracy is achieved through the image registration function of digital pathology software.

[0107] In step S204, the spatial association is a unified spatial coordinate system for multimodal data. This system is constructed based on the H&E staining structure of pathological images, and a feature point matching algorithm maps the second spatial transcriptome data and the second single-cell RNA sequencing data to the same reference system. This step preserves the spatial accuracy of the original data, allowing for joint analysis of data from different modalities at the same spatial scale.

[0108] In step S205, batch effect correction preferably employs a negative binomial distribution model to identify technical variations, and the ComBat algorithm eliminates the influence of sample preparation and sequencing batches. The corrected second-space transcriptome data retains the true biological differences, ensuring the reliability of cross-sample comparisons.

[0109] In step S206, cell integrity criteria include indicators such as the number of detected genes and the proportion of ribosomal genes. High-quality cells are selected through automated cluster analysis combined with manual verification. The filtered second single-cell RNA sequencing data improves the accuracy of downstream analysis and avoids interference from low-quality cells.

[0110] In step S207, the quality control database can be constructed using a clinical research data management system. Preprocessed multi-omics data is stored in HDF5 format and associated with metadata indexes, supporting standardized management and traceability query of multi-center research data.

[0111] This embodiment achieves precise integration of spatial transcriptomics, single-cell sequencing, and pathological image data through a systematic multimodal data quality control and alignment process. High-quality spatial loci are screened based on technical parameters and biological characteristics, anatomical regions are precisely matched through affine transformation, and image registration ensures multimodal spatial alignment. Batch correction and cell filtering eliminate technical variations, ultimately constructing a standardized dataset in a unified coordinate system. This embodiment combines histological localization accuracy with molecular feature analysis, ensuring that spatial multi-omics data retains its original biological characteristics while maintaining comparability. This provides high-fidelity input for subsequent cross-scale analysis, improving the reliability and analytical depth of multimodal data integration.

[0112] Please see Figure 3 In some embodiments, cross-modal semantic embedding of second spatial transcriptome data is performed based on second single-cell RNA sequencing data to generate spatially enhanced expression profiles, including:

[0113] S301. Extract the cell type feature matrix from the second single-cell RNA sequencing data. The cell type feature matrix contains the expression profile of the marker genes of each cell type.

[0114] S302. Align each spatial site of the second spatial transcriptome data with the cell type feature matrix by comparison and learning, and calculate the similarity score between the spatial site and each cell type.

[0115] S303. Based on similarity scores, perform probabilistic cell type deconvolution on spatial sites to generate a cell type ratio matrix. The deconvolution is jointly optimized using nonnegative matrix factorization and Markov random field.

[0116] S304. Sparse signal completion is performed on low-expression genes in the second space transcriptome data. The completion signal is derived from the prior expression distribution of the same type of cells in single-cell data.

[0117] S305. By fusing the cell type ratio matrix with the completed gene expression data, a spatially enhanced expression profile is generated. The spatially enhanced expression profile includes cell composition information and gene expression confidence scores for each spatial site.

[0118] In step S301, the cell type feature matrix is ​​a molecular feature map of cell subpopulations obtained through single-cell clustering analysis. Its construction process includes: normalizing the second single-cell RNA sequencing data using standard single-cell analysis procedures such as Seurat; identifying major cell subpopulations through PCA dimensionality reduction and UMAP visualization; and finally extracting differentially expressed marker genes for each cell type to form the feature matrix. In the cell type feature matrix, row vectors represent genes, column vectors represent cell types, and matrix elements represent the average expression level of marker genes in a specific cell type.

[0119] In step S302, the contrastive learning alignment is achieved by learning the shared feature space between single cells and spatial data through a neural network model. Specifically, the maximum mean difference is used as a measure of distribution distance to project the mixed expression profile of spatial sites onto the single-cell feature space. The similarity score is calculated using cosine similarity to reflect the degree of matching between the expression pattern of spatial sites and the reference profile of each cell type. The score calculation takes into account the dropout noise characteristics unique to spatial transcriptomics.

[0120] In step S303, the joint optimization process of nonnegative matrix factorization (NMF) and Markov random field (MRF) can be understood as follows: Cell type proportions are initially estimated using NMF; then, an MRF model is introduced to integrate cell composition similarity constraints between adjacent spatial sites. Its energy function includes a data fitting term and a spatial smoothing term. Preferably, the optimization process employs a variational expectation-maximization algorithm, where the spatial smoothing weights are adaptively adjusted based on the continuity of the tissue region.

[0121] In step S304, cells of the same type possess conserved gene expression programs, providing a basis for sparse signal completion of low-expression genes. For genes not detected in the spatial data, completion values ​​are sampled from the expression distribution of their respective cell types in single-cell data. Preferably, the distribution parameters are estimated using Bayesian methods, considering covariates such as gene length and sequencing depth. Furthermore, the completed expression values ​​are accompanied by a confidence score, reflecting the expression stability of the gene in the corresponding cell type.

[0122] In step S305, the fusion process preferably employs a weighted splicing method. The cell type ratio matrix provides information on the spatial microenvironment composition, and the completed gene expression data retains the original detection signal. The two are correlated through spatial coordinates. In the generated spatially enhanced expression profile, the gene expression confidence score comprehensively considers the quality of the original UMI count, the reliability of the completed signal, and the cell type specificity.

[0123] This embodiment establishes cross-modal semantic correspondences to transfer the fine classification capability at the single-cell level to spatial data while preserving the original spatial topological information. Probabilistic deconvolution handles cell type co-localization, and sparse completion restores biological signals lost due to technical reasons. The final output spatially enhanced expression profile combines the classification accuracy of single-cell resolution with the in-situ preservation advantage of spatial transcriptomics, making it particularly suitable for the analysis of highly heterogeneous tissues such as the tumor microenvironment. It can accurately resolve the complex spatial interaction patterns between malignant cells and stromal cells.

[0124] Please see Figure 4 In some embodiments, probabilistic cell type deconvolution is performed on spatial sites based on similarity scores to generate a cell type ratio matrix. The deconvolution is jointly optimized using nonnegative matrix factorization and Markov random fields, including:

[0125] S401. Construct a non-negative matrix factorization objective function of spatial site-cell type using similarity scores as initial weights. Establish a neighborhood graph of Markov random field based on the spatial coordinate information of the second spatial transcriptome data. The neighborhood graph contains the topological connection relationship between spatial sites.

[0126] S402. Introduce a spatial smoothing constraint term into the non-negative matrix factorization objective function. The spatial smoothing constraint term adjusts the cell type ratio difference between adjacent sites through a Markov random field.

[0127] S403. The cell type proportion matrix is ​​iteratively updated using an alternating optimization algorithm. Each iteration includes:

[0128] With fixed space constraints, the cell type ratio of the non-negative matrix factorization is optimized by multiplication update rules; and with fixed cell type ratios, spatial consistency is optimized by graph Laplacian regularization.

[0129] The optimization terminates when the change in the Frobenius norm of the cell type ratio matrix in adjacent iterations is less than the preset convergence threshold.

[0130] The optimized cell type ratio matrix is ​​normalized to ensure that the sum of the cell type ratios at each spatial location is 1;

[0131] Output a cell type ratio matrix. The rows of the cell type ratio matrix correspond to spatial locations, the columns correspond to cell types, and the element values ​​are probabilistic ratio weights.

[0132] In step S401, the nonnegative matrix factorization objective function is used to decompose the mixed expression profile of spatial sites into a linear combination of cell type feature matrix and proportion matrix. The similarity score, as an initial weight, effectively guides the decomposition direction. Utilizing the cross-modal correspondence between single-cell data and spatial data, the decomposition results are more consistent with biological reality. The neighborhood graph of the Markov random field is constructed using spatial coordinate information. Preferably, the Delaunay triangulation algorithm is used to determine the topological connections between sites, which can adaptively reflect the spatial organization structure of regions with different densities, providing accurate neighborhood definitions for subsequent spatial constraints.

[0133] In step S402, the spatial smoothing constraint term is introduced into the objective function in the form of graph Laplacian regularization. Spatial continuity is maintained by adjusting the degree of difference in the proportion of cell types between adjacent sites. The constraint strength parameter is dynamically adjusted according to the degree of heterogeneity of the tissue region, which avoids boundary blurring caused by excessive smoothing and effectively suppresses spurious differences caused by technical noise. The potential function design of the Markov random field takes into account the biological characteristics of cell type co-localization, enabling the model to distinguish between real cell mixing regions and artificial artifacts.

[0134] In step S403, the alternating optimization algorithm gradually approaches the optimal solution through a decomposition and iteration process. The multiplicative update rule ensures the non-negativity of the cell type ratio, while graph Laplacian regularization efficiently achieves spatial consistency constraints through matrix operations. The convergence threshold is set by comprehensively considering the balance between analytical accuracy and computational efficiency. The final result after normalization meets the basic requirements of the probability distribution. Furthermore, a temperature coefficient is introduced to adjust and preserve information on minor cellular components in the microenvironment.

[0135] This embodiment organically combines the molecular precision of single-cell data analysis with the topological information of spatial transcriptomics. The construction of adaptive neighborhood relationships and the strength of dynamic constraints enable the model to adapt to both the clear boundaries of the tumor core region and the gradual characteristics of the infiltration region. This provides a reliable computational framework for the quantitative analysis of complex tissue microenvironments and is particularly suitable for application scenarios that require accurate spatial composition information, such as tumor immunotherapy response prediction.

[0136] In some embodiments, multi-scale mapping is performed on the second spatial transcriptome data and the second pathological image data to extract spatial heterogeneity features, including:

[0137] Based on single-cell resolution expression profiles from second-space transcriptome data, a micro-cell interaction network diagram was constructed. This micro-cell interaction network diagram was generated through ligand-receptor pair co-expression analysis and cell adjacency calculation.

[0138] Tissue region segmentation was performed on the second pathological image data to identify the tumor core area, invasion front area and mesenchymal area. Combined with the expression characteristics of the second spatial transcriptome data of the corresponding regions, a mesoscopic functional region map was constructed.

[0139] By integrating the tissue structure features of second pathological image data at the whole-slice level with the global expression patterns of spatial transcriptome data, a macroscopic tissue metastasis trend prediction map was established.

[0140] Extract tumor-immune cell interaction strength features from microscopic cell interaction network diagrams, including the spatial colocalization frequency of immune checkpoint molecular pairs.

[0141] The spatial gradient characteristics of the vascular invasion front were calculated from the mesoscopic functional region atlas, including the attenuation coefficient of endothelial marker expression with distance.

[0142] Topological features of high-risk sub-regions were extracted from the macro-organization migration trend prediction map, including the spatial autocorrelation index of the microenvironment before migration.

[0143] The characteristics of tumor-immune cell interaction intensity, spatial gradient of vascular invasion front, and topological characteristics of high-risk subregions are integrated into spatial heterogeneity features.

[0144] In this embodiment, the microscopic cell-cell interaction network map is a cell-cell interaction relationship map constructed based on single-cell resolution expression profiling. It is generated by analyzing the co-expression patterns of ligand-receptor pairs and combining them with cell spatial adjacency relationships to characterize the molecular communication features between cells in the tumor microenvironment. The ligand-receptor pair co-expression analysis uses a validated cell-cell interaction database as a reference, and cell adjacency relationships are determined by dividing the spatial coordinates using a Voronoi diagram. This multimodal integration method can accurately identify biologically significant cell interaction events.

[0145] Mesoscopic functional region atlases were obtained by segmenting tissue regions from pathological images. The tumor core region was identified based on cell density and nuclear atypia characteristics, the invasion front region was delineated based on the morphological characteristics of the tumor-stromal junction, and the stromal region was defined by collagen fiber distribution and the degree of immune cell infiltration. The expression characteristics of the second spatial transcriptome data of each functional region were integrated by spatial coordinate alignment.

[0146] The macroscopic tissue metastasis trend prediction map integrates whole-slice-level histopathological features with global spatial transcriptomic expression patterns. Tissue structural features include morphological indicators such as the degree of vascular invasion and the distribution of necrotic areas, while the global expression pattern is extracted after dimensionality reduction through principal component analysis. The macroscopic tissue metastasis trend prediction map uses graph neural networks to model the spatial propagation patterns of tissue structures to infer potential pathways of tumor metastasis.

[0147] The intensity characteristics of tumor-immune cell interactions were extracted from the microscopic network diagram, focusing on the spatial co-localization of immune checkpoint molecular pairs. The calculation process considered the weighted relationship between the product of ligand-receptor expression levels and intercellular distance. The spatial gradient characteristics of the vascular invasion front were obtained by fitting the decay curves of endothelial marker expression as a function of distance, with the decay coefficient calculated using a nonlinear regression method. The spatial autocorrelation index in the topological characteristics of high-risk subregions was quantified using Moran's I statistic, reflecting the degree of regional homogeneity of the pre-metastatic microenvironment.

[0148] This embodiment constructs a complete spatial heterogeneity analysis framework by extracting and fusing features at the micro, meso, and macro scales. The micro-level cell interaction network map reveals cell interaction mechanisms, the meso-level functional region map locates functional region characteristics, and the macro-level tissue metastasis trend prediction map grasps tissue evolution trends. This multi-scale integrated approach can comprehensively analyze the spatial dynamics of tumor development. In terms of immunotherapy response prediction, by integrating immune checkpoint co-localization features and vascular invasion patterns, more precise spatial basis can be provided for treatment plan selection.

[0149] In some embodiments, the tissue structure features of second pathological image data at the whole-slice level are integrated with the global expression patterns of spatial transcriptome data to establish a macroscopic tissue metastasis trend prediction map, including:

[0150] Full-slice scanning was performed on the second pathological image data to extract tissue structural features, including the orientation of tumor glands, the density of stromal fibrosis distribution, and the spatial proportion of necrotic areas.

[0151] Spatial autocorrelation analysis was performed on the global expression patterns of the second spatial transcriptome data to calculate the spatial clustering index of expression hotspots of each gene. The global expression patterns included the spatial distribution of expression levels of epithelial-mesenchymal transition markers, angiogenesis factors, and immunosuppression-related molecules.

[0152] A multimodal fusion model is constructed based on graph convolutional neural networks. It aligns tissue structure features with global expression patterns. The alignment process uses a cross-modal attention mechanism to allocate weights between pathological image features and transcriptome expression features, and outputs a fused multimodal feature matrix.

[0153] A macroscopic tissue metastasis trend prediction map is established in the multimodal feature matrix. The macroscopic tissue metastasis trend prediction map simulates the potential metastasis path of tumor cells along tissue structure features and expression gradient features through a random walk algorithm.

[0154] Topology optimization is performed on the macro-organization transfer trend prediction map to identify high-risk transfer sub-regions. Topology optimization includes calculating the spatial density of transfer path intersections and the consistency coefficient of transfer direction.

[0155] The final output is a macro-organizational migration trend prediction map, which includes a migration probability heatmap, high-risk sub-region boundary markings, and a vector field of major migration paths.

[0156] In this embodiment, the directional arrangement of tumor glands refers to the distribution characteristics of the main axis direction of glandular structures identified through digital pathological image analysis. Preferably, a gray-level co-occurrence matrix combined with a directional filter is used to calculate and reflect the invasive growth pattern of tumor tissue. The distribution density of stromal fibrosis can be quantitatively analyzed in Masson staining regions using a deep learning segmentation model (such as U-Net), further excluding non-specific staining interference from blood vessel walls, etc. The spatial proportion of necrotic areas is determined by morphological processing combined with threshold segmentation. It should be noted that consistency verification with manual interpretation results is performed to ensure clinical relevance.

[0157] In spatial autocorrelation analysis, gene expression hotspots can be identified using spatial statistical methods. The spatial clustering index of epithelial-mesenchymal transition markers is calculated using the local Moran's I algorithm. For the expression gradient characteristics of angiogenic factors, thin-plate spline interpolation is recommended for smoothing. The distribution pattern analysis of immunosuppressive molecules needs to consider the truncation effect of tissue boundaries. Global expression pattern analysis should maintain data consistency with spatially enhanced expression profiling.

[0158] Preferably, the cross-modal attention mechanism is implemented based on the Transformer architecture, where the query vector comes from pathological image features, the key-value pairs come from transcriptome features, and the weight allocation process needs to incorporate relative position encoding to maintain spatial correspondence. Furthermore, the generation of the multimodal feature matrix can employ graph convolution operations, and the construction of the adjacency matrix should be clearly distinguished from the scale of the microscopic cellular interaction network graph.

[0159] When simulating potential metastatic pathways of tumor cells along tissue structural features and expression gradients using the random walk algorithm, it is important to note the following: the anisotropic migration weights provided by the glandular arrangement direction, the resistance coefficient derived from fibrosis density, and the absolute prohibition constraint of necrotic regions. Preferably, the walk step size should be parameterized according to the histological scale, and the simulation results need to be independently repeated multiple times to ensure stability.

[0160] In the process of topology optimization of the macroscopic tissue metastasis trend prediction map, the calculation of the metastasis direction consistency coefficient excludes the interference of local eddy regions, and the determination of the high-risk sub-region boundary is calibrated in conjunction with the pathologically defined tumor front region. The final vector field visualization can use a streamline tracing algorithm while maintaining precise correspondence with the original spatial coordinate system.

[0161] This embodiment constructs a multi-scale, interconnected framework for predicting tumor metastasis. Pathological features such as the directional arrangement of tumor glands and the density of stromal fibrosis distribution are fused with molecular expression patterns such as epithelial-mesenchymal transition markers and angiogenesis factors through a graph convolutional neural network. Cross-modal attention mechanisms are used to align features, and a random walk algorithm simulates the migration paths of tumor cells along tissue structures and expression gradients. This embodiment establishes a predictive bridge from microscopic molecular features to macroscopic tissue morphology. The output macroscopic tissue metastasis trend prediction map includes a metastasis probability heatmap, high-risk subregion boundary annotations, and vector fields of major metastasis pathways. This provides clinical assessment of tumor metastasis risk with both spatial resolution and molecular specificity, helping to guide surgical boundary planning and targeted therapy strategy development. It is particularly suitable for assessing the metastasis risk of highly heterogeneous tumors and guiding precise intervention.

[0162] In some embodiments, spatially enhanced expression profiles and spatial heterogeneity features are input into a pre-trained metastasis risk prediction model, which outputs a spatial heatmap containing liver metastasis probabilities and a list of key driving features, including:

[0163] Feature normalization was performed on the spatially enhanced expression profile to make the expression levels of different genes comparable, including Z-score normalization and batch effect correction.

[0164] Spatial heterogeneity features are dimensionally compressed, and principal component analysis is used to retain principal component features whose contribution rate exceeds a preset threshold.

[0165] The standardized spatial augmented expression spectrum and the compressed spatial heterogeneous features are concatenated to generate a fused feature matrix;

[0166] The fused feature matrix is ​​input into a pre-trained graph neural network model, which aggregates multi-scale spatial information based on an attention mechanism.

[0167] The probability of liver metastasis at each spatial site is calculated using a graph neural network model, and a spatial heatmap is generated based on the probability value. The spatial heatmap maintains a spatial correspondence with the original tissue slice.

[0168] The gradient backpropagation method is used to extract key driving features from the graph neural network model, including gene expression features and spatial topology features that contribute the most to the prediction results;

[0169] The key driving features are sorted according to their contribution, and a list of key driving features including feature name, contribution score and biological explanation is generated.

[0170] The spatial heatmap is associated with and stored in a list of key driving features, and a mapping relationship with clinicopathological parameters is established.

[0171] In this embodiment, the standardization of spatially enhanced expression profiles employs Z-score normalization to eliminate differences in gene expression dimensions, and uses the ComBat algorithm for batch effect correction to ensure comparability between different samples. Furthermore, tissue-region-specific correction factors can be introduced during the standardization process, employing differentiated normalization strategies for different microenvironmental regions such as the tumor core and invasion front, to better preserve biologically relevant spatial expression variations.

[0172] Spatial heterogeneity characteristics characterize the spatial variation patterns of gene expression in the tissue microenvironment. Dimensional compression is achieved through principal component analysis, with a preferred preset threshold set to principal component features contributing 85% cumulatively. This threshold can be determined using the elbow rule or parallel analysis. Preferably, when retaining principal components, spatial autocorrelation analysis results can be incorporated, prioritizing feature components exhibiting significant spatial clustering patterns.

[0173] The fused feature matrix generated by feature splicing is used to integrate molecular expression profiles and spatial distribution features. The preferred implementation method is to perform horizontal splicing along the feature dimensions while maintaining the correspondence of the original spatial coordinates. Optionally, mutual information analysis can be performed on the two types of features before splicing, and the splicing weights can be dynamically adjusted based on the correlation between features to avoid interference from redundant information.

[0174] The pre-trained graph neural network model achieves multi-scale spatial information aggregation based on a multi-head attention mechanism. The graph structure construction considers both spatial proximity and feature similarity connections, and the node feature update process employs a graph attention network (GAT) architecture. Furthermore, the attention mechanism can introduce a spatial distance decay factor, causing the information transmission weights of neighboring nodes to decrease exponentially with increasing distance, which better aligns with biological diffusion patterns.

[0175] The spatial heatmap is generated by mapping the predicted probability value of each spatial location to the corresponding tissue region, and the spatial correspondence is ensured by registration with H&E stained sections. Preferably, histological boundary information can be superimposed during heatmap generation, so that the prediction results and pathological morphological features can be intuitively correlated.

[0176] When extracting key driving features using the gradient backpropagation method, the integral gradient method is employed to calculate feature contribution, focusing on identifying gene expression features and spatial topological features that have the greatest impact on liver metastasis prediction. Optionally, the contribution calculation can be validated using Shapley values ​​to ensure the stability of feature importance assessment.

[0177] Biological interpretation of the key driver feature list can be achieved through KEGG pathway enrichment analysis and spatial colocalization validation, and its mapping relationship with clinicopathological parameters is established using Spearman rank correlation coefficient calculation. Furthermore, a feature-phenotype association network can be established to visualize the complex interactions between key driver genes and clinical indicators. Preferably, the spatial heatmap and feature list are stored using a spatial database structure, supporting joint queries based on location and features, and can be expanded to integrate electronic medical record data to achieve multidimensional association analysis.

[0178] This embodiment achieves multi-scale information fusion from molecular expression features to tissue spatial heterogeneity by concatenating standardized spatially enhanced expression profiles with spatially heterogeneous features compressed by principal component analysis into a fusion feature matrix, which is then input into an attention-based graph neural network model. The spatial heatmap generated by this method maintains an accurate correspondence with the original tissue slices, while the key driver feature list extracted through gradient backpropagation reveals the molecular mechanisms of liver metastasis. The predictive model established in this embodiment retains the molecular specificity of spatial transcriptome data while integrating the spatial heterogeneity features of the tissue microenvironment. The output spatial heatmap and driver feature list provide clinicians with decision-making support that combines spatial resolution and molecular interpretability, making it particularly suitable for assessing liver metastasis risk and guiding individualized treatment.

[0179] In a second aspect, the present invention also provides a spatial omics-based colorectal cancer metastasis prediction device, applicable to the method described in the first aspect. The device includes a data acquisition module, a data processing module, a logic operation module, and a report generation module. The data acquisition module is used to acquire raw multi-omics data from original colorectal cancer tissue samples. The raw multi-omics data includes first spatial transcriptome data, first single-cell RNA sequencing data, and first pathological image data.

[0180] The data processing module is used to perform modality alignment and quality control on the raw multi-omics data to obtain preprocessed multi-omics data, which includes second spatial transcriptome data, second single-cell RNA sequencing data, and second pathological image data.

[0181] The logic operation module is used to perform cross-modal semantic embedding on the second spatial transcriptome data based on the second single-cell RNA sequencing data, generating a spatially enhanced expression profile. The cross-modal semantic embedding includes single-cell-spatial feature alignment and probabilistic sparse signal completion based on contrastive learning. The spatially enhanced expression profile includes cell type deconvolution results and confidence scores of metastasis-related genes. Multi-scale graph construction is performed on the second spatial transcriptome data and the second pathological image data to extract spatial heterogeneity features. Multi-scale graph construction includes microscopic cell interaction graphs, mesoscopic functional region graphs, and hierarchical modeling of macroscopic tissue structures. Spatial heterogeneity features include tumor-immune adjacency strength, spatial gradient of the vascular invasion front, and topological relationships of high-risk subregions. The spatially enhanced expression profile and spatial heterogeneity features are input into a pre-trained metastasis risk prediction model, and the output includes a spatial heatmap containing the probability of liver metastasis and a list of key driving features.

[0182] The report generation module is used to generate a clinical prediction report on the risk of colorectal cancer metastasis based on spatial heat maps and a list of key driving features. The report includes the location of high-risk areas, metastasis probability scores, and treatment response predictions.

[0183] In this embodiment, the data acquisition module can be a hardware system integrating a high-throughput sequencer and a digital pathology scanner, used to automatically collect raw multi-omics data of colorectal cancer tissue samples, and it ensures data input compatibility through a standardized interface.

[0184] The data processing module is deployed on high-performance computing nodes and is equipped with modality alignment algorithms and quality control pipelines to generate preprocessed multi-omics data that can be directly used for analysis. Preferably, the data processing module accelerates large-scale data processing through a parallel computing architecture.

[0185] The logic operation module is preferably implemented using a GPU-accelerated server. Its core is a pre-trained graph neural network model, which, through built-in cross-modal embedding units and a multi-scale graph analysis engine, transforms the input spatially enhanced expression spectrum and spatial heterogeneity features into liver metastasis risk prediction results. The output interface of the logic operation module supports standardized data exchange formats.

[0186] The report generation module integrates visualization tools and clinical report templates to automatically combine the spatial heatmaps output by the model with a list of key driving features into an interactive HTML report, which can be directly accessed by the electronic medical record system of medical institutions.

[0187] This embodiment automates the entire process of predicting colorectal cancer liver metastasis through modular design. The data acquisition module ensures standardized acquisition of multi-omics data, the data processing module achieves seamless integration of multimodal information, the logic operation module uses hardware acceleration to efficiently analyze complex spatial features, and the report generation module provides clinically friendly decision support output. This device transforms cutting-edge spatial omics analysis methods into a standardized, deployable medical device solution, significantly improving the reliability and clinical applicability of colorectal cancer metastasis risk assessment, and providing pathology departments with end-to-end technical support from experimental data to treatment recommendations.

[0188] In a third aspect, this embodiment also provides a computer-readable storage medium storing computer program instructions thereon, which, when executed by a processor, implement the method described in the first aspect.

[0189] The computer program involved in this embodiment can be stored in a computer device readable storage medium, which includes, but is not limited to, disks, magnetic tapes, magnetic cards, floppy disks, flash memory, optical disks, optical cards, read-only memory (ROM), random access memory (RAM), erasable programmable ROM (EPROM), and electrically erasable programmable ROM (EEPROM), etc. It also includes other biological, physical, or chemical structures capable of performing similar or equivalent functions to the storage media listed above, such as DNA, RNA, proteins, and other units with information storage capabilities. In specific embodiments, the storage medium involved can be one of the above-mentioned media types or a combination of the above media types. In different embodiments, the computer program involved in the embodiment can be centrally stored in a single medium or distributed across multiple media. The memory containing the computer device readable storage medium can be non-volatile memory or random access memory. These computer device readable storage media can be built into the device or connected to the device involved in the embodiment as an external device or part of an external device. In some embodiments, the memory having a computer device readable storage medium is deployed locally; in other embodiments, the memory may be deployed remotely from the processor, for example, as a network-attached memory accessed via RF circuitry or an external port and a communication network, wherein the communication network may be the Internet, one or more intranets, a local area network (LAN), a wide area network (WLAN), a storage area network (SAN), or a suitable combination thereof, as long as computer device access to the memory is enabled. Furthermore, the computer program involved in the embodiments may be stored in plaintext / ciphertext form, or it may be designed as training data, integrated and recombined through model training and implicitly stored in the parameter states of a deep neural network or other machine learning model.

[0190] Please see Figure 5 In a fourth aspect, this embodiment also provides an electronic device 1, including a memory 11 and a processor 12, wherein the memory 11 is used to store one or more computer program instructions, wherein the one or more computer program instructions are executed by the processor 12 to implement the method described in the first aspect.

[0191] The processor described in this embodiment can be implemented by hardware, firmware, software, or a combination thereof. It can be a circuit, one or more of an application-specific integrated circuit (ASIC), a digital signal processor (DSP), a digital signal processing device (DSPD), a programmable logic device (PLD), a field-programmable gate array (FPGA), a central processing unit (CPU), a controller, a microcontroller, or a microprocessor. It also includes other physical, biological, or chemical structures that can implement the same or equivalent functions as the processors listed above, such as biological neurons, quantum computing units, DNA computing units, etc., so that the processor can execute some or all of the steps in the computer program or method involved in the various embodiments of this application, or any combination of the steps mentioned therein.

[0192] Unlike existing technologies, the above-mentioned technical solution offers the following advantages: By integrating spatial transcriptome data, single-cell RNA sequencing data, and pathological image data, a multi-scale spatial omics analysis system is constructed. Cross-modal semantic embedding technology is used to fuse single-cell resolution features with spatial transcriptome data, generating a spatially enhanced expression profile that includes cell type deconvolution results. Through hierarchical modeling of microscopic cell interaction maps, mesoscopic functional region maps, and macroscopic tissue structure maps, spatial heterogeneity features including tumor-immune adjacency strength, spatial gradient of the vascular invasion front, and topological relationships of high-risk subregions are extracted. Finally, based on graph neural networks, molecular expression and spatial topological features are fused to output a spatial heatmap of liver metastasis probability and a list of key driving features. Ultimately, a clinical prediction report is generated, including high-risk region localization, metastasis probability scoring, and treatment response prediction. This technical solution achieves multi-dimensional correlation analysis from gene expression to tissue morphology, providing a single-cell precision and spatial resolution solution for colorectal cancer liver metastasis risk assessment. It has advantages in predictive accuracy and clinical applicability, and is valuable for guiding individualized treatment decisions.

[0193] Finally, it should be noted that although the above embodiments have been described in the text and drawings of this application, this should not limit the scope of patent protection of this application. Any technical solutions that are based on the essential concept of this application and utilize the content described in the text and drawings of this application, resulting in equivalent structural or procedural substitutions or modifications, as well as the direct or indirect application of the technical solutions of the above embodiments to other related technical fields, are all included within the scope of patent protection of this application.

Claims

1. A device for predicting colorectal cancer metastasis based on spatial omics, characterized in that, The device includes: The data acquisition module is used to acquire raw multi-omics data from raw colorectal cancer tissue samples, including first spatial transcriptome data, first single-cell RNA sequencing data, and first pathological image data. The data processing module is used to perform modality alignment and quality control processing on the raw multi-omics data to obtain preprocessed multi-omics data, which includes second spatial transcriptome data, second single-cell RNA sequencing data, and second pathological image data. The logic operation module is used to perform cross-modal semantic embedding on the second spatial transcriptome data based on the second single-cell RNA sequencing data to generate a spatially enhanced expression profile. The cross-modal semantic embedding includes single-cell-spatial feature alignment and probabilistic sparse signal completion based on contrastive learning. The spatially enhanced expression profile includes cell type deconvolution results and confidence scores of metastasis-related genes. The module also performs multi-scale graph construction on the second spatial transcriptome data and the second pathological image data to extract spatial heterogeneity features. The multi-scale graph construction includes hierarchical modeling of microscopic cell interaction maps, mesoscopic functional region maps, and macroscopic tissue structures. The spatial heterogeneity features include tumor-immune adjacency strength, spatial gradient of the vascular invasion front, and topological relationships of high-risk subregions. Finally, the module inputs the spatially enhanced expression profile and spatial heterogeneity features into a pre-trained metastasis risk prediction model, outputting a spatial heatmap containing liver metastasis probability and a list of key driving features. The report generation module is used to generate a clinical prediction report on the risk of colorectal cancer metastasis based on the spatial heatmap and the list of key driving features. The report includes the location of high-risk areas, metastasis probability score, and treatment response prediction. Based on the second single-cell RNA sequencing data, cross-modal semantic embedding is performed on the second spatial transcriptome data to generate a spatially enhanced expression profile, including: A cell type feature matrix is ​​extracted from the second single-cell RNA sequencing data, the cell type feature matrix containing the expression profile of marker genes for each cell type; Each spatial site of the second spatial transcriptome data is compared and aligned with the cell type feature matrix, and the similarity score between the spatial site and each cell type is calculated. Based on the similarity score, probabilistic cell type deconvolution is performed on the spatial sites to generate a cell type ratio matrix. The deconvolution is jointly optimized by nonnegative matrix factorization and Markov random field. Sparse signal completion was performed on the low-expression genes in the second spatial transcriptome data. The completion signal was derived from the prior expression distribution of the same type of cells in single-cell data. By fusing the cell type ratio matrix with the completed gene expression data, a spatially enhanced expression profile is generated, which includes cell composition information and gene expression confidence scores for each spatial locus. Multi-scale maps were constructed from the second spatial transcriptome data and the second pathological image data to extract spatial heterogeneity features, including: Based on the single-cell resolution expression profile of the second spatial transcriptome data, a microscopic cell interaction network diagram is constructed. The microscopic cell interaction network diagram is generated through ligand-receptor pair co-expression analysis and cell adjacency relationship calculation. The second pathological image data is segmented into tissue regions to identify the tumor core area, invasion front area and stroma area. Combined with the expression characteristics of the second spatial transcriptome data of the corresponding regions, a mesoscopic functional region map is constructed. By integrating the tissue structure features of second pathological image data at the whole-slice level with the global expression patterns of spatial transcriptome data, a macroscopic tissue metastasis trend prediction map was established. The intensity characteristics of tumor-immune cell interactions were extracted from the microscopic cell interaction network diagram, including the spatial colocalization frequency of immune checkpoint molecular pairs. The spatial gradient characteristics of the vascular invasion front were calculated from the mesoscopic functional region atlas, including the attenuation coefficient of endothelial marker expression as a function of distance. The topological features of high-risk sub-regions, including the spatial autocorrelation index of the microenvironment before migration, are extracted from the macro-organization migration trend prediction map. The tumor-immune cell interaction intensity characteristics, vascular invasion front spatial gradient characteristics, and high-risk subregion topological characteristics are fused into spatial heterogeneity characteristics.

2. The colorectal cancer metastasis prediction device based on spatial omics according to claim 1, characterized in that, The original multi-omics data undergoes modality alignment and quality control processing to obtain preprocessed multi-omics data, including: Spatial sites that meet a preset quality threshold are selected from the first spatial transcriptome data as the second spatial transcriptome data. The preset quality threshold includes a gene detection threshold and an upper limit for the proportion of mitochondrial genes. The cell sequencing data that matches the anatomical region of the second spatial transcriptome data in the first single-cell RNA sequencing data are extracted as the second single-cell RNA sequencing data. The anatomical region matching is achieved by tissue slice coordinate mapping. The image region corresponding to the acquisition location of the second spatial transcriptome data is selected from the first pathological image data as the second pathological image data, and the spatial deviation of the acquisition location during matching does not exceed the preset alignment tolerance. Establish spatial correlations among the second spatial transcriptome data, the second single-cell RNA sequencing data, and the second pathological image data, so that the three have a unified coordinate reference system; Batch effect correction was performed on the second spatial transcriptome data to eliminate technical variations between different samples; The second single-cell RNA sequencing data was filtered for low-quality cells, retaining transcriptome data that met the cell integrity standard. Generate preprocessed multi-omics data, and store multiple preprocessed multi-omics data in a quality control database.

3. The colorectal cancer metastasis prediction device based on spatial omics according to claim 1, characterized in that, Based on the similarity score, probabilistic cell type deconvolution is performed on spatial sites to generate a cell type ratio matrix. The deconvolution is jointly optimized using nonnegative matrix factorization and Markov random fields, including: The similarity score is used as the initial weight to construct a non-negative matrix factorization objective function for spatial sites and cell types. A neighborhood graph of a Markov random field is established based on the spatial coordinate information of the second spatial transcriptome data. The neighborhood graph contains the topological connections between spatial sites. A spatial smoothing constraint term is introduced into the nonnegative matrix factorization objective function. The spatial smoothing constraint term adjusts the cell type ratio difference between adjacent sites through a Markov random field. The cell type proportion matrix is ​​updated iteratively using an alternating optimization algorithm. Each iteration includes: With fixed space constraints, the cell type ratio of the non-negative matrix factorization is optimized by multiplication update rules; and with fixed cell type ratios, spatial consistency is optimized by graph Laplacian regularization. The optimization terminates when the change in the Frobenius norm of the cell type ratio matrix in adjacent iterations is less than the preset convergence threshold. The optimized cell type ratio matrix is ​​normalized to ensure that the sum of the cell type ratios at each spatial location is 1; Output the cell type ratio matrix, where rows correspond to spatial locations, columns correspond to cell types, and element values ​​are probabilistic ratio weights.

4. The colorectal cancer metastasis prediction device based on spatial omics according to claim 1, characterized in that, By integrating the structural features of second pathological image data at the whole-slice level with the global expression patterns of spatial transcriptome data, a macroscopic tissue metastasis trend prediction map was established, including: The second pathological image data was scanned in its entirety to extract tissue structural features, including the directional arrangement of tumor glands, the density of stromal fibrosis distribution, and the spatial proportion of necrotic areas. Spatial autocorrelation analysis was performed on the global expression patterns of the second spatial transcriptome data to calculate the spatial clustering index of expression hotspot regions of each gene. The global expression patterns include the spatial distribution of expression levels of epithelial-mesenchymal transition markers, angiogenesis factors, and immunosuppression-related molecules. A multimodal fusion model is constructed based on graph convolutional neural networks. The tissue structure features are aligned with the global expression patterns. The alignment process uses a cross-modal attention mechanism to allocate the weights of pathological image features and transcriptome expression features, and outputs a fused multimodal feature matrix. A macroscopic tissue metastasis trend prediction map is established in the multimodal feature matrix. The macroscopic tissue metastasis trend prediction map simulates the potential metastasis path of tumor cells along tissue structure features and expression gradient features through a random walk algorithm. Topology optimization is performed on the macro-organizational migration trend prediction map to identify high-risk migration sub-regions. The topology optimization includes calculating the spatial density of migration path intersections and the consistency coefficient of migration direction. The final macro-organizational migration trend prediction map is output, which includes a migration probability heatmap, high-risk sub-region boundary markings, and a vector field of major migration paths.

5. The colorectal cancer metastasis prediction device based on spatial omics according to claim 1, characterized in that, The spatially enhanced expression profile and spatial heterogeneity features are input into a pre-trained metastasis risk prediction model, which outputs a spatial heatmap containing liver metastasis probability and a list of key driving features, including: The spatially enhanced expression profile is subjected to feature normalization to make the expression levels of different genes comparable, including Z-score normalization and batch effect correction. The spatial heterogeneity features are dimensionally compressed, and principal component analysis is used to retain principal component features whose contribution rate exceeds a preset threshold. The standardized spatial augmented expression spectrum and the compressed spatial heterogeneous features are concatenated to generate a fused feature matrix; The fused feature matrix is ​​input into a pre-trained graph neural network model, which aggregates multi-scale spatial information based on an attention mechanism. The probability of liver metastasis at each spatial site is calculated using the graph neural network model, and a spatial heatmap is generated based on the probability value. The spatial heatmap maintains a spatial correspondence with the original tissue slice. Key driving features, including gene expression features and spatial topology features that contribute the most to the prediction results, are extracted from the graph neural network model using the gradient backpropagation method. The key driving features are sorted according to their contribution to generate a list of key driving features that includes feature name, contribution score and biological explanation. The spatial heatmap is associated with and stored in a list of key driving features, and a mapping relationship with clinical pathological parameters is established.

Images