Gastric cancer multi-omics marker detection method, system and equipment

Abstract

The invention discloses a gastric cancer multi-omics marker detection method, system and device, and the method comprises the following steps: S1, collecting a public database open-source space transcriptome, a single cell sequencing sample and bulk-RNAseq data for pre-processing, and collecting a primary tissue sample of a gastric cancer patient in the center for data processing; s2, integrating different modal data samples to obtain a patient label of an input end, and constructing a marker detection model and training the marker detection model; and S3, extracting a target feature value from the input external pathological section by using the trained model, and generating a diagnosis prediction result. Through multi-modal data chimerism, algorithm optimization and AI system development, subpopulation cell marker proportion prediction and prognosis diagnosis and marker evaluation with population prognosis information are realized, and an integrated diagnosis scheme for breaking through molecule-space-prognosis information is constructed.

Description

Technical Field

[0001] This invention relates to the field of medical information technology, and in particular to a method, system and device for detecting multi-omics biomarkers of gastric cancer. Background Technology

[0002] Peritoneal metastasis of gastric cancer is a leading cause of death in patients with advanced gastric cancer, and its early diagnosis is crucial for improving survival rates. However, current clinical diagnosis and research face three major technical bottlenecks: insufficient sensitivity of clinical diagnostic methods, leading to missed diagnoses or misdiagnoses; data and technology gaps in multi-omics biomarker research: the proportion of gastric cancer peritoneal metastasis samples in public databases relying on spatial transcriptomics is <0.1%; in single-cell transcriptomics analysis, existing protocols do not integrate spatial location information and cannot analyze cancer cell invasion patterns; furthermore, the level of intelligence in pathological diagnosis is low, traditional manual slide reading cannot quantify the spatial distribution density of specific cell subpopulations, and diagnostic reproducibility is poor; existing AI models only analyze image texture and do not integrate molecular biomarkers. In summary, due to data silos, technological fragmentation, and clinical disconnect, the discovery and application efficiency of gastric cancer peritoneal metastasis biomarkers is low. Summary of the Invention

[0003] The purpose of this invention is to provide a method, system, and device for detecting multi-omics biomarkers in gastric cancer, in order to solve the problems mentioned in the background art.

[0004] To achieve the above-mentioned objectives, one aspect of the present invention provides a method for detecting multi-omics biomarkers in gastric cancer, comprising the following steps:

[0005] Step S1: Collect open-source spatial transcriptome, single-cell sequencing samples, and bulk-RNAseq data from public databases for preprocessing; collect primary tissue samples from gastric cancer patients at our center for data processing.

[0006] Step S2: Integrate data samples from different modalities to obtain patient labels at the input end, construct a biomarker detection model, and train it.

[0007] Step S3: Using the trained model, extract target feature values ​​from the input external pathological slices to generate diagnostic prediction results.

[0008] Furthermore, step S2 includes the following steps:

[0009] Step S21, Pre-identification of spatial transcriptome sample biomarkers;

[0010] Step S22: Annotate spatial slices with single-cell precision and search for markers;

[0011] Step S23: Optimize the assessment of prognostic biomarkers in batch RNA-seq;

[0012] Step S24: Label the pathological images and train the biomarker detection model;

[0013] Furthermore, step S21 includes the following steps:

[0014] Step S211: Using a tumor boundary estimation tool and based on the pathological features annotated by a pathologist, the spatial region is divided into a tumor core (TC), an invasive margin (IM), and a non-tumor (NT) region, and each transcriptomic region is mapped to the corresponding pathological region.

[0015] Step S212: Perform differential gene expression analysis on the tumor core area, invasion border area, and non-tumor area;

[0016] Step S213: Perform unsupervised cluster analysis on the tumor core area, invasion edge area, and non-tumor area respectively;

[0017] Step S214: Using a spatial neighborhood model, calculate the co-occurrence probability of a specific cell type or gene expression pattern in physical space.

[0018] Furthermore, step S22 includes the following steps:

[0019] Step S221: Obtain a high-quality scRNA-seq dataset of gastric cancer for standard analysis, which includes dimensionality reduction and clustering, and provides fine cell type annotation for each cell cluster;

[0020] Step S222: Ensure that the spatial data and the gene set covered by the single-cell data are consistent, and perform necessary batch effect corrections.

[0021] Step S223: The selected statistical model / deconvolution method models the mixed expression profile of a spatial region as a weighted sum of the reference expression profiles of various cell types contained therein;

[0022] Step S224: Map individual cells or cell types from scRNA-seq space to the most similar location in the spatial transcriptome;

[0023] Step S225: Run the selected algorithm to transfer or deconvolve the cell type tags and state information defined by scRNA-seq onto each region of the spatial transcriptome;

[0024] Step S226: Extract the number of cells of each cell type in each sample and calculate the proportion. Perform statistical tests or use a mixed-effects model to test the differences between groups in order to analyze and compare the differences in the proportion of each type of cell in different sample groups.

[0025] Furthermore, step S23 includes the following steps:

[0026] Step S231: Screen for subpopulation-specific genes and generate a feature gene matrix, where rows are genes and columns are target cell subpopulations;

[0027] Step S232: Download the transcriptome data of gastric cancer from TCGA (The Cancer Genome Atlas) and the corresponding clinical data, and perform preprocessing.

[0028] Step S233: Select the deconvolution method and infer the cell proportion;

[0029] Step S234: Integrate the deconvolution results with clinical data;

[0030] Step S235: Group the proportions of each cell type separately and perform survival analysis.

[0031] Furthermore, step S24 includes the following steps:

[0032] Step S241: Obtain pathological image data of patients in the TCGA queue;

[0033] Step S242: Preprocess the pathological image data, including segmentation and filtering;

[0034] Step S243: Incorporate labeled patient image patches into the training model;

[0035] Step S244: Sample image patches and divide them into training set, validation set and test set;

[0036] Step S245: Use ResNet50 as the base model for deep learning artificial intelligence image training;

[0037] Step S246: Evaluate the model's performance on the test set.

[0038] Furthermore, the training process in step S245 is as follows: replace the last fully connected layer of ResNet50 so that its output has two nodes, corresponding to high risk and low risk respectively; use the cross-entropy loss function; use the optimizer for training, set the initial learning rate to 0.001, and use the learning rate scheduling strategy; use the validation set to monitor the model performance during training, and save the best model.

[0039] Furthermore, data augmentation techniques are applied to the graphic blocks to increase data diversity and improve the model's generalization ability. These data augmentation techniques include rotation, flipping, and color dithering.

[0040] A second aspect of the present invention provides a multi-omics biomarker detection system for gastric cancer, comprising an acquisition module, a sample integration and analysis module, and a diagnostic module, wherein:

[0041] The acquisition module is used to collect open-source spatial transcriptome, single-cell sequencing samples, and bulk-RNAseq data from public databases for preprocessing, and to collect primary tissue samples from gastric cancer patients in our own center for data processing.

[0042] The sample integration and analysis module is used to integrate data samples from different modalities, obtain patient labels at the input end, construct biomarker detection models, and train them.

[0043] The diagnostic module is used to extract target feature values ​​from input external pathological slices using the trained model and generate diagnostic prediction results.

[0044] A third aspect of the present invention provides an electronic device comprising:

[0045] One or more processors;

[0046] Memory;

[0047] and one or more programs stored in memory, the one or more programs including instructions for performing the gastric cancer multi-omics biomarker detection method as described above.

[0048] Compared with existing technologies, this system and method have the following advantages:

[0049] 1. This invention can systematically expand the peritoneal metastasis-specific dataset by chimerizing public space transcriptome samples and self-test samples, thus solving the sample bias problem;

[0050] 2. This invention utilizes deep learning algorithms to concatenate single-cell transcriptome data with histological images to achieve high-resolution capture of gene expression heterogeneity;

[0051] 3. In the validation of prognostic biomarkers, this invention uses BayesPrism and CIBERSORT optimized reference matrices, combined with self-tested bulk RNA-seq data and public cohorts, to achieve biomarker evaluation with population prognostic information.

[0052] 4. This invention combines the expression characteristics of specific cell populations screened by multi-omics with TCGA pathological slide data to train a ResNet50 deep learning model, and develops a detection system and electronic device that can input external slides to achieve prediction of the proportion of subpopulation cell markers and prognostic diagnosis.

[0053] 5. This invention constructs an integrated diagnostic scheme that connects molecular, spatial, and prognostic information through multimodal data integration, algorithm optimization, and AI system development. Attached Figure Description

[0054] Figure 1 This is a flowchart of a multi-omics biomarker detection method for gastric cancer.

[0055] Figure 2 This is a schematic diagram of a multi-omics biomarker detection system for gastric cancer.

[0056] Figure 3 Survival curves for prognostic risk assessment using the present invention.

[0057] Figure 4 This is a comparison chart of the detection results output of the present invention and the prior art. Detailed Implementation

[0058] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0059] like Figure 1 The diagram shown is a flowchart of a method according to one embodiment. This embodiment of the invention provides a multimodal fusion method for detecting multi-omics biomarkers in gastric cancer, and the specific steps are as follows:

[0060] Step S1: Obtain raw samples and collect open-source spatial transcriptome, single-cell sequencing samples, and bulk-RNAseq data from public databases for preprocessing, including batch removal, normalization, and quality control. Next, collect primary tissue samples, tumor tissues, and peritoneal metastasis samples from gastric cancer patients gathered at our center for batch removal, normalization, and quality control.

[0061] Step S2: The different types of data that have been collected and preprocessed are analyzed and integrated step by step according to the following four steps S21 to S24.

[0062] Step S21: Pre-identification of spatial transcriptome sample biomarkers. Using SpaGCN software combined with histological images, the spatial transcriptome data is divided into three spatial domains: tumor core region, invasive border region, and non-tumor region. Differentially expressed genes (DEGs) in each region are identified, and spatial spots in different regions are clustered. This includes the following steps:

[0063] In step S211, after obtaining the quality-controlled and batch-corrected spatial transcriptome data, we need to use tumor boundary estimation tools and histopathological features annotated by pathologists to divide the spatial spots into tumor core areas, invasive border areas, and non-tumor areas. Then, through spatial coordinate registration, each transcriptome spot is mapped to the corresponding pathological region. For example, spots located in areas with dense tumor cells are labeled TC, spots located at the junction of tumor and normal tissue and exhibiting tumor cell infiltration characteristics are labeled IM, and areas composed entirely of normal tissue structures are labeled NT.

[0064] In step S212, we performed differentially expressed gene analysis on these three functional regions. A generalized linear model based on a negative binomial distribution was used to perform pairwise comparisons of TC vs IM, TC vs NT, and IM vs NT. The screening criteria for differentially expressed genes were typically set as an absolute logarithmic fold change (|log2FC|) greater than 0.5 and a false discovery rate (FDR) less than 0.05. Through this analysis, we can identify region-specific marker genes. For example, the tumor core may highly express cell cycle-related genes (such as MKI67), the invasive periphery may be enriched with epithelial-mesenchymal transition (EMT) genes (such as VIM), while non-tumor regions highly express tissue-specific functional genes (such as ALB in the liver).

[0065] Step S213: Building upon the regional segmentation and differential expression analysis, we further explore the heterogeneity within each region. This requires unsupervised clustering analysis for each of the three regions. The specific process includes: first, principal component analysis (PCA) dimensionality reduction using region-specific hypervariable genes; then, constructing a K-nearest neighbor graph based on the principal component space; and finally, identifying subpopulations using the Leiden algorithm. For example, in the tumor core region, hypoxia response subpopulations (highly expressing HIF1A) and proliferation subpopulations (highly expressing TOP2A) may be clustered, while in the invasive border region, subpopulations enriched with immune cells may be identified. The spatial distribution patterns of these subpopulations can be visualized using t-SNE or UMAP, and their tissue localization characteristics can be displayed by overlaying spatial locations.

[0066] Step S214: To gain a deeper understanding of spatial interactions between cells, we also need to perform co-location analysis. Here, a spatial neighborhood model is used to calculate the co-occurrence probability of specific cell types or gene expression patterns in physical space. Commonly used statistical methods include the cross-type L-function or the co-location index (CI).

[0067] Step S22 involves single-cell precision annotation of spatial slices and biomarker identification. This involves using the STAMapper software tool to transfer cell type tags from single-cell RNA sequencing data to single-cell spatial transcriptomics data, achieving high-precision cell type mapping and annotation. Differential analysis is then used to analyze the high-precision cell subpopulations, identifying specific cell subpopulations as biomarkers.

[0068] One limitation of spatial transcriptomics is that each spot captures a mixture of RNA from multiple cells (averaging a few to dozens of cells). To resolve the specific cell type composition and state within each spot, it is necessary to integrate high-resolution single-cell RNA sequencing data. This step is called cell type deconvolution or spatial mapping. It includes the following steps:

[0069] Step S221: Obtain the high-quality scRNA-seq dataset for gastric cancer that matches the research samples from Step S1. Perform standard analyses on this integrated dataset: dimensionality reduction, clustering, and fine-grained cell type annotation for each cell cluster (e.g., CD4+ Naive T, CD8+ Exhausted T, M2 Macrophage, Myofibroblast, Tumor Epithelial Cell Clone A, etc.). This well-annotated scRNA-seq dataset will serve as a "reference map".

[0070] Step S222: Ensure that the spatial data and the gene set covered by the single-cell data are consistent, and perform the necessary batch effect correction (even if it has been done before, fine-tuning may be required during integration).

[0071] Step S223, selecting the integration / mapping algorithm, is the most crucial step. We chose a statistical model / deconvolution-based method: Stereoscope and RCTD (Robust Cell Type Decomposition). This involves modeling the mixed expression profile of a spatial spot as a weighted sum of the reference expression profiles (from scRNA-seq) of the various cell types it contains. The algorithm solves for the proportion or absolute abundance of each cell type in each spot through optimization (such as nonnegative matrix factorization (NMF) and Bayesian inference). Cell2location is increasingly popular because it can estimate absolute cell density.

[0072] Step S224 uses similarity / mapping methods, such as Seurat's CCA + Label Transfer. It focuses on "mapping" individual cells or cell types from the scRNA-seq space to the most similar location in the spatial transcriptome. Tangram is particularly adept at mapping while maintaining spatial continuity.

[0073] Step S225: Perform integration and annotation by running the selected algorithm to “transfer” or “deconvolve” the cell type tags and state information defined by scRNA-seq to each spot of the spatial transcriptome.

[0074] Step S226: Then extract the number of cells of each cell type in each sample and calculate the proportion. Perform statistical tests (optional), or use a mixed-effects model (considering intra-sample correlation) to test inter-group differences in order to analyze and compare the differences in the proportion of each type of cell in different sample groups (normal, primary lesion, metastatic lesion).

[0075] Step S23 involves optimizing the technology for bulk RNA-seq prognostic biomarker assessment. Based on BayesPrism and CIBERSORT deconvolution software, the reference matrix is ​​optimized to cover gastric cancer metastasis-specific cell populations. Bulk RNA-seq data is deconvolved to infer the proportion of each cell type in the tissue. A biomarker assessment model is constructed by combining population prognostic information (such as survival rate and recurrence rate). This includes the following steps:

[0076] Step S231: After completing single-cell data analysis and identifying key endothelial and fibroblast subpopulations, we need to extract the molecular characteristics of these subpopulations as "fingerprints." This is typically done by calculating differentially expressed genes for each subpopulation. Using functions similar to FindAllMarkers, we set thresholds (e.g., logFC > 0.5, p < 0.01) to screen for subpopulation-specific genes. Special attention is paid to retaining housekeeping genes that are stably expressed in multiple subpopulations as internal controls. The final result is a feature gene matrix, where rows represent genes (approximately 300-500 key genes), columns represent target cell subpopulations (endothelial subpopulation 1, endothelial subpopulation 2, fibroblast subpopulation A, etc.), and matrix values ​​represent the average expression level of genes in that subpopulation. This matrix will serve as a reference blueprint for deconvolution.

[0077] Step S232: TCGA data comes from tumor tissue samples of gastric cancer patients, and its transcriptome data is the expression profile of the entire tissue (containing mixed signals from multiple cell types). Next, we need to download the TCGA gastric cancer transcriptome data (usually an RNA-seq gene expression matrix) and corresponding clinical data (including survival time, survival status, etc.). Data preprocessing is performed: gene names are standardized (e.g., HGNC symbol), low-expression genes are filtered (TPM>1 in >50% of samples), and logarithmic transformation (Log2(TPM+1)). Special attention is paid to removing batch effects, and the ComBat algorithm is used to correct for differences between sequencing centers.

[0078] Step S233: Select the deconvolution method and infer cell proportions: The purpose of deconvolution is to decompose the expression profile of TCGA samples into the contribution proportions of each cell type using a reference matrix constructed from single-cell data. Commonly used tools include CIBERSORTx, MuSiC, and EPIC. Here, we chose CIBERSORTx: Upload the single-cell reference matrix to the local CIBERSORTx platform. Set the deconvolution parameters: Select "deconvolution mode" and specify the reference matrix. Use the TCGA expression matrix as mixed input data and run the deconvolution algorithm. The algorithm will output a cell proportion matrix for each TCGA sample, including the estimated proportions of endothelial cells and fibroblasts in each sample (and the proportions of other background cell types, even if more types are included in the reference matrix, but here we only focus on these two target subpopulations).

[0079] Step S234, Integrate deconvolution results with clinical data: After obtaining the estimated proportions of the two cell types in the TCGA samples, merge them with the TCGA clinical data (survival data). Download or organize the clinical data from TCGA, including the patient survival time (e.g., OS: overall survival time) and survival status (e.g., whether the patient has died) for each sample. Match the cell proportion matrix obtained from deconvolution with the clinical data using the sample ID.

[0080] Step S235, Survival Analysis: Group the samples by the proportion of each cell type (endothelial cells and fibroblasts). A common method is to divide the samples into "high proportion group" and "low proportion group" according to the median or optimal cutoff point (determined by the survminer package). Visualize the survival differences using Kaplan-Meier curves: the horizontal axis is time, and the vertical axis is survival rate; plot two curves (high proportion group and low proportion group).

[0081] Step S24 involves labeling and pre-training pathological images. The expression of specific cell populations discovered by the above method is grouped into high and low groups within the population. This is combined with existing pathological slide information from the TCGA database to perform deep learning AI image training based on ResNet50 software. This includes the following steps:

[0082] Step S241: Based on the previous sub-procedure, we already have the cell proportion deconvolution score for patients in the TCGA cohort. According to this score, we can divide patients into high-expression and low-expression groups. Simultaneously, we acquire the pathological image data of these patients (usually H&E-stained whole-slide images, WSIs). It is important to note that each patient may have multiple pathological images, and we need to ensure that the images are correctly correlated with the patient's risk score.

[0083] Step S242: The pathological images provided by TCGA are ultra-high resolution WSI images (typically tens of thousands of pixels by tens of thousands of pixels), which cannot be directly input into a deep learning model. Therefore, we need to segment each WSI image into small image patches, 224x224 pixels (this is the standard input size for ResNet50). Segmentation typically uses a sliding window approach, and the stride can be set to control the degree of overlap. Furthermore, we filter out disorganized areas (such as blank areas, tissue folds, or artifact areas), which can be achieved using tissue segmentation algorithms.

[0084] In step S243, since we have risk labels at the patient level, all image patches for each patient will inherit that patient's risk label. However, a challenge is that a slice may contain both tumor and normal regions simultaneously, and we unsupervisedly incorporate these into the training model.

[0085] Step S244: Divide the image patches obtained in the previous step into High and Low groups according to the patient. Since each patient may generate thousands of image patches, and the entire dataset may be very large, we need to perform appropriate sampling to avoid data imbalance and computational resource issues. Then, divide all image patches into training, validation, and test sets. Note that during partitioning, it is necessary to ensure that image patches from the same patient do not appear in both the training and validation / test sets simultaneously to prevent data leakage.

[0086] like Figure 3The diagram shows the Kaplan-Meier risk assessment survival curve using the technical solution of this invention. According to the method mentioned in this invention, gastric cancer patients are divided into intervals. Specifically, T represents the identified cell type A; M represents the identified cell type B; H represents the population with high expression of this cell type; and L represents the population with low expression of this cell type. Based on the above classification method, the population containing both type A and type B cells is divided into four categories.

[0087] Figure 3 The survival curves shown have the horizontal axis representing the overall survival time range of patients and the vertical axis representing the survival probability. The graph shows that the high-risk group differentiated by this invention has the worst survival prognosis, with a p-value less than 0.001, indicating the reliability of the current invention's prognostic value.

[0088] In step S245, we choose ResNet50 as the base model. Since pathological images differ from natural images (ImageNet), we typically use weights pre-trained on ImageNet for transfer learning. The specific training process is as follows:

[0089] Replace the last fully connected layer of the ResNet50 to output two nodes (corresponding to high-risk and low-risk). Use the cross-entropy loss function. Train using an optimizer (such as SGD or Adam) with a small initial learning rate (e.g., 0.001) and a learning rate scheduling strategy (e.g., decay on demand). Monitor model performance using a validation set during training and save the best model. To fully utilize image patches, we can use data augmentation techniques (such as rotation, flipping, color jittering, etc.) to increase data diversity and improve the model's generalization ability.

[0090] Step S246: Evaluate the model's performance on the test set. Since our labels are at the patient level, while the model makes predictions at the image patch level, we need to integrate the predictions from multiple image patches for each patient to obtain patient-level predictions. Integration methods can include majority voting or averaging prediction probabilities (and then selecting the class with the higher probability). Then, calculate metrics such as patient-level accuracy and AUC. Furthermore, we can use visualization techniques such as Grad-CAM to understand the regions the model focuses on and verify whether it is focused on the pathological features of the tumor.

[0091] Step S3 is to execute the image evaluation strategy:

[0092] The model trained by the above-mentioned integration-training module is packaged into a deployable module. For new pathological images (exogenous images), the same preprocessing is first performed (segmenting into image patches and filtering out non-tissue regions and possibly selecting only tumor regions). Then, the trained model is used to predict for each image patch, and finally, the prediction results of all image patches are integrated (again using voting or average probability) to obtain the diagnostic efficacy prediction at the patient level. In this way, we can input pathological images of new patients into the system and automatically output risk classifications.

[0093] like Figure 2 The diagram shown is a system structure diagram of another embodiment of the present invention. This system includes an acquisition module, a sample integration and analysis module, and a diagnostic module, wherein:

[0094] The acquisition module is used to collect open-source spatial transcriptome, single-cell sequencing samples, and bulk-RNAseq data from public databases for preprocessing, and to collect primary tissue samples from gastric cancer patients in our own center for data processing.

[0095] The sample integration and analysis module is used to integrate data samples from different modalities, obtain patient labels at the input end, construct biomarker detection models, and train them; it includes four sub-modules.

[0096] The first fusion submodule executes the spatial transcriptome sample biomarker pre-identification analysis strategy.

[0097] The second fusion submodule performs single-cell precision annotation of spatial slices and implements biomarker search analysis strategies.

[0098] The third fusion submodule: Optimization analysis strategy for performing batch RNA-seq prognostic biomarker evaluation.

[0099] The fourth fusion submodule: performs pathological image labeling and pre-trains strategies.

[0100] The diagnostic module is used to extract target feature values ​​from input external pathological slices using the trained model and generate diagnostic prediction results.

[0101] like Figure 4As shown, the right image is the output of the detection results using this invention, while the left image is the result of detection based on existing empty transcriptome sequencing data. Specifically, the left image shows the detection results after spatial transcriptome sequencing of a pathological slide. Red dots represent highly expressed cells of interest in the sequencing data, and red areas represent the regions where highly expressed cells of interest are present. Spatial transcriptome sequencing is very expensive and is currently an advanced digital method for detecting cells in pathological slides. The right image shows the detection tool developed by this invention, which is highly efficient and accurate. Specifically, the right image is a slide heatmap; darker areas represent more detected cells of interest. The horizontal and vertical axes of the right image represent spatial locations without specific numerical meaning. Comparing the two images, it can be seen that the tool developed by this invention may rival spatial transcriptome sequencing and identification methods in terms of precision and accuracy, and this invention is more efficient.

[0102] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for detecting multi-omics biomarkers in gastric cancer, characterized in that, Includes the following steps: Step S1: Collect open-source spatial transcriptome, single-cell sequencing samples, and bulk-RNAseq data from public databases for preprocessing; collect primary tissue samples from gastric cancer patients at our center for data processing. Step S2: Integrate data samples from different modalities to obtain patient labels at the input end, construct a biomarker detection model, and train it. Step S3: Using the trained model, extract target feature values ​​from the input external pathological slices to generate diagnostic prediction results.

2. The method for detecting multi-omics biomarkers of gastric cancer according to claim 1, characterized in that, Step S2 includes the following steps: Step S21, Pre-identification of spatial transcriptome sample biomarkers; Step S22: Annotate spatial slices with single-cell precision and search for markers; Step S23: Optimize the assessment of prognostic biomarkers in batch RNA-seq; Step S24: Label the pathological images and train the biomarker detection model.

3. The method for detecting multi-omics biomarkers of gastric cancer according to claim 2, characterized in that, Step S21 includes the following steps: Step S211: Using a tumor boundary estimation tool and based on the pathological features annotated by a pathologist, the spatial region is divided into a tumor core region, an invasion edge region, and a non-tumor region, and each transcriptomic region is mapped to the corresponding pathological region. Step S212: Perform differential gene expression analysis on the tumor core area, invasion border area, and non-tumor area; Step S213: Perform unsupervised cluster analysis on the tumor core area, invasion edge area, and non-tumor area respectively; Step S214: Using a spatial neighborhood model, calculate the co-occurrence probability of a specific cell type or gene expression pattern in physical space.

4. The method for detecting multi-omics biomarkers of gastric cancer according to claim 2, characterized in that, Step S22 includes the following steps: Step S221: Obtain a high-quality scRNA-seq dataset of gastric cancer for standard analysis, which includes dimensionality reduction and clustering, and provides fine cell type annotation for each cell cluster; Step S222: Ensure that the spatial data and the gene set covered by the single-cell data are consistent, and perform necessary batch effect corrections; Step S223: The selected statistical model / deconvolution method models the mixed expression profile of a spatial region as a weighted sum of the reference expression profiles of various cell types contained therein; Step S224: Map individual cells or cell types from scRNA-seq space to the most similar location in the spatial transcriptome; Step S225: Run the selected algorithm to transfer or deconvolve the cell type tags and state information defined by scRNA-seq onto each region of the spatial transcriptome; Step S226: Extract the number of cells of each cell type in each sample and calculate the proportion. Perform statistical tests or use a mixed-effects model to test the differences between groups in order to analyze and compare the differences in the proportion of each type of cell in different sample groups.

5. The method for detecting multi-omics biomarkers of gastric cancer according to claim 2, characterized in that, Step S23 includes the following steps: Step S231: Screen for subpopulation-specific genes and generate a feature gene matrix, where rows are genes and columns are target cell subpopulations; Step S232: Download the transcriptome data of gastric cancer from TCGA and the corresponding clinical data, and perform preprocessing. Step S233: Select the deconvolution method and infer the cell proportion; Step S234: Integrate the deconvolution results with clinical data; Step S235: Group the proportions of each cell type separately and perform survival analysis.

6. The method for detecting multi-omics biomarkers of gastric cancer according to claim 2, characterized in that, Step S24 includes the following steps: Step S241: Obtain pathological image data of patients in the TCGA queue; Step S242: Preprocess the pathological image data, including segmentation and filtering; Step S243: Incorporate labeled patient image patches into the training model; Step S244: Sample image patches and divide them into training set, validation set and test set; Step S245: Use ResNet50 as the base model for deep learning artificial intelligence image training; Step S246: Evaluate the model's performance on the test set.

7. The method for detecting multi-omics biomarkers of gastric cancer according to claim 6, characterized in that, The training process in step S245 is as follows: Replace the last fully connected layer of ResNet50 so that its output has two nodes, corresponding to high risk and low risk respectively; use the cross-entropy loss function; use the optimizer for training, set the initial learning rate to 0.001, and use the learning rate scheduling strategy. During training, a validation set is used to monitor model performance, and the best model is saved.

8. The method for detecting multi-omics biomarkers of gastric cancer according to claim 6, characterized in that, Data augmentation techniques are applied to graphic blocks to increase data diversity and improve the generalization ability of the model. These data augmentation techniques include rotation, flipping, and color dithering.

9. A multi-omics biomarker detection system for gastric cancer, characterized in that, It includes an acquisition module, a sample integration and analysis module, and a diagnostic module, among which: The acquisition module is used to collect open-source spatial transcriptome, single-cell sequencing samples, and bulk-RNAseq data from public databases for preprocessing, and to collect primary tissue samples from gastric cancer patients in our own center for data processing. The sample integration and analysis module is used to integrate data samples from different modalities, obtain patient labels at the input end, construct biomarker detection models, and train them. The diagnostic module is used to extract target feature values ​​from input external pathological slices using the trained model and generate diagnostic prediction results.

10. An electronic device, characterized in that, include: One or more processors; Memory; and one or more programs stored in memory, the one or more programs including instructions for performing the gastric cancer multi-omics biomarker detection method as described in any one of claims 1-8.

Images