Method, device and system for constructing a disease single cell as detection marker panel

Abstract

The application discloses a disease single cell AS detection marker panel construction method, device and system, and belongs to the field of database construction. Compared with a conventional database, the obtained marker panel is used for sequencing samples with different geographical sources on a plateau and a plain, so that dynamic changes of variable splicing driven by regional factors can be revealed, and defects of single data dimension and large deviation from a real scene of the conventional database can be made up. Secondly, cell type labels, barcodes and PSI matrices are integrated to construct a high-resolution variable splicing map, and cell heterogeneity and space-time expression patterns can be accurately captured. Finally, the AS marker panel obtained through differential screening and function annotation can provide more accurate, reliable and comprehensive research data for researchers.

Description

Technical Field

[0001] This application relates to the field of database construction technology, and in particular, to a method, apparatus and system for constructing a panel of biomarkers for single-cell AS detection. Background Technology

[0002] Disease research is currently shifting from static "genetic blueprints" to dynamic "life scripts," and transcriptomics is the core tool for deciphering this script. By capturing the dynamics and spatiotemporal changes of gene expression within cells in a panoramic way, it allows us to glimpse the deep molecular logic of disease occurrence and development.

[0003] However, current transcriptomics research uses conventional databases, which have limited data dimensions and may even differ significantly from reality. Summary of the Invention

[0004] To overcome the shortcomings of related technologies, this application provides a method, device, and system for constructing a panel of biomarkers for single-cell AS detection of diseases, in order to solve the problem that the databases used in existing disease transcriptomics research are of a single dimension and differ significantly from the actual situation.

[0005] The technical solution adopted by this application to solve its technical problem is: Firstly, a method for constructing a panel of biomarkers for single-cell AS detection is provided, including: Single-cell transcriptome sequencing was performed to obtain sequencing files, which recorded the geographical origin of each sample, including plains and plateau regions; The sequencing file is processed to obtain target data, which includes: type label for each cell, cell barcode, and sorting BAM file for each sample; The sorted BAM file was input into various tools to obtain the PSI matrix; A variable splicing map is constructed based on the geographic origin, the cell type label, the cell barcode, and the PSI matrix; The variable splice map is subjected to differential AS marker screening and functional annotation to obtain an AS detection marker panel.

[0006] Further, the processing of the sequencing file to obtain the target data includes: Evaluation metrics are calculated based on the sequencing files. Data that does not meet the preset requirements for evaluation metrics are removed to obtain high-quality data. The evaluation metrics include Q30 base ratio, GC content distribution, and adapter contamination ratio. For the high-quality data, intermediate data is obtained by comparing the genomes using a preset human genome as a reference genome. The intermediate data includes: a raw cell × gene count matrix and a sorted BAM file containing cell barcodes and UMI information. The quality control indicators of the intermediate data are obtained, and low-quality cells are removed based on the quality control indicators. The quality control indicators include the number of genes detected per cell, the total number of UMIs per cell, and the proportion of mitochondrial genes. The data for removing low-quality cells is used to predict and remove cells predicted to be dichotomous. The target data is obtained by batch integration and cell type annotation of the data after removing duplex cells. The batch integration and cell type annotation include batch effect correction, dimensionality reduction and clustering, and cell type annotation.

[0007] Furthermore, the step of inputting the sorted BAM file into various tools to obtain the PSI matrix includes: Obtain the output of multiple tools separately; The outputs of the multiple tools are integrated to obtain the original PSI matrix, where rows represent AS events and columns represent cells; The original PSI matrix that meets the following filtering conditions is retained to obtain the final PSI matrix, wherein the filtering conditions include: the cell coverage of each AS event is greater than or equal to the preset cell coverage, the PSI standard deviation of each AS event across all cells is greater than or equal to the preset standard deviation, and the number of valid AS events detected per cell is greater than or equal to the preset number of events.

[0008] Furthermore, the process of integrating the outputs of the multiple tools to obtain the original PSI matrix includes: The AS event coordinates in the output of the multi-tool are uniformly converted into a standard format, and the standard format includes the AS event type in the AS event coordinates. For AS events of type SE, A5SS, A3SS, or MXE, the event is retained only if a valid PSI value is detected in at least two preset tools; for AS events of type IR, the output of one preset tool is taken as the primary factor, and the event is retained after the output of the other preset tool is verified.

[0009] Further, the construction of the variable splicing map based on the geographic origin, the cell type label, the cell barcode, and the PSI matrix includes: The PSI matrix is ​​matched and merged with the cell type label and the geographic origin, and the cell barcode is used as the unique connection key to construct a complete PSI annotation matrix containing geographic hierarchical information. The mean cell type-specific PSI was calculated for each AS event, using cell type and geographic origin as the unit. Within each cell type, the differences in AS between plateau and plain regions were compared, and the PSI difference was calculated. A variable splicing map is obtained based on the complete PSI annotation matrix, the specific PSI mean, and the PSI difference.

[0010] Further, the step of performing differential AS marker screening and functional annotation on the variable splice map to obtain an AS detection marker panel includes: Within each cell type, the PSI difference between plateau and plain regions was calculated, and the corrected P value was also calculated. Functional annotation is performed only on AS events whose absolute PSI difference is greater than or equal to a preset PSI difference and whose corrected P value is less than a preset corrected P value. The functional annotation includes: gene-level annotation, protein structure-level annotation, immune function-level annotation, and TB-specific association annotation. For each candidate AS biomarker that passes functional annotation, a comprehensive priority score is calculated based on the importance of its biological function. For each cell type, the biomarkers are sorted in descending order of comprehensive priority score, and the top K1 high-confidence AS biomarkers are retained to obtain the AS detection biomarker panel, where K1 is a positive integer.

[0011] Further, the step of performing differential AS marker screening and functional annotation on the variable splice map to obtain an AS detection marker panel includes: A three-layer heterogeneous graph was constructed using AS events, host genes, and splicing factors. The node set contained three types of nodes: AS event nodes with PSI matrix as node feature vectors; host gene nodes with gene expression level as features; and splicing factor nodes with expression level and known binding motif information as features. The edge set contained three types of edges: weighted co-expression edges based on gene-gene co-expression correlation coefficients; biological prior edges based on known pathway relationships from the GO / KEGG pathway database; and regulatory edges based on known regulatory relationships between splicing factors and their target AS events. A graph attention network is used to learn a neighbor-weighted aggregation representation for each AS event node. The attention weights reflect the contributions of different neighboring nodes to the importance of the current AS event. After multi-layer graph convolution aggregation, a low-dimensional embedding vector is generated for each AS event. The embedding vector simultaneously encodes the expression features of the AS event itself and its topological location information in the control network. After completing graph convolution feature learning, interpretability attribution methods are further integrated to quantitatively clarify the contribution of each AS event node to the final classification task, and to extract key interaction relationships between AS events and between AS events and splicing factors. By combining the mean attention weight and gradient attribution score of each AS event node, the final feature importance comprehensive score is calculated. All AS events are sorted in descending order of the final feature importance comprehensive score, and the top K2 high-confidence AS events are retained to obtain the AS detection marker panel, where K2 is a positive integer.

[0012] Further, the step of performing differential AS marker screening and functional annotation on the variable splice map to obtain an AS detection marker panel includes: Using the PSI matrix of sample × AS event as input, each AS event is treated as an independent token in the sequence, and the initial embedding vector of each AS event is constructed by combining learnable position encoding with AS event type encoding. The initial embedding vector is input into the Transformer encoder, and the attention weights between any two AS events are calculated through a multi-head self-attention mechanism to capture their joint dependency structure. The weight matrix of each attention head is extracted. The comprehensive importance score of each AS event across all samples and all attention heads is calculated by cross-head mean aggregation and hierarchical attribution method. AS events that are significantly higher than the background level are selected into the candidate panel pool according to the importance score. The Transformer model is fine-tuned using downstream supervision tasks on labeled samples to obtain the fine-tuned attention weights. The candidate panel pool is filtered based on the attention weights to obtain the AS detection marker panel.

[0013] Secondly, a device for constructing a panel of biomarkers for single-cell AS disease detection is provided, comprising: The sequencing file acquisition module is used to perform single-cell transcriptome sequencing to obtain sequencing files. The sequencing files record the geographical origin of each sample, including plains and plateaus. The target data acquisition module is used to process the sequencing file to obtain target data, which includes: type label for each cell, cell barcode, and sorting BAM file for each sample; The PSI matrix acquisition module is used to input the sorted BAM file into various tools to obtain the PSI matrix. The AS map construction module is used to construct a variable splice map based on the geographic origin, the cell type label, the cell barcode, and the PSI matrix; The marker panel construction module is used to perform differential AS marker screening and functional annotation on the variable splice map to obtain an AS detection marker panel.

[0014] Thirdly, a panel construction system for single-cell AS disease detection biomarkers is provided, including: At least one processor and at least one memory; The memory stores the executable instructions of the processor; The processor is configured to execute the method for constructing a panel of biomarkers for single-cell AS detection provided in the first aspect of the technical solution.

[0015] Beneficial effects

[0016] This application provides a method, apparatus, and system for constructing a biomarker panel for single-cell AS detection. Compared to conventional databases, the constructed biomarker panel, by sequencing samples from different geographical origins (plateau and plains), can reveal regionally driven dynamic changes in alternative splicing, overcoming the shortcomings of conventional databases, such as limited data dimensionality and significant deviation from real-world scenarios. Secondly, by integrating cell type tags, barcodes, and PSI matrices, a high-resolution alternative splicing atlas is constructed, accurately capturing cellular heterogeneity and spatiotemporal expression patterns. Finally, the AS biomarker panel, obtained through differential screening and functional annotation, can provide researchers with more accurate, reliable, and comprehensive research data. Attached Figure Description

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

[0018] Figure 1 This is a flowchart of a method for constructing a panel of biomarkers for single-cell AS detection provided in an embodiment of this application; Figure 2 This is a schematic diagram of a panel construction device for single-cell AS detection biomarkers provided in an embodiment of this application. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of this application clearer, the technical solutions of this application will be described in detail below with reference to the accompanying drawings and embodiments. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments of this application, all other implementation methods obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0020] Definition: PSI (Percent Spliced ​​In): Measures the proportion of a particular exon that is incorporated into mature mRNA, with a value ranging from 0 to 1. PSI=1 indicates that the exon is incorporated into all transcripts; PSI=0 indicates complete skipping.

[0021] Marker Panel: Analogous to the Antibody Panel in flow cytometry, the Marker Panel in this application is a set of validated AS events that can be used to standardize the annotation of cell types and disease states in new samples.

[0022] Existing transcriptome studies for infectious diseases (such as tuberculosis and sepsis) mostly rely solely on gene expression levels to analyze the state of immune cells. However, alternative splicing events can independently regulate protein function and immune responses even when gene expression levels are not significantly different. In the absence of single-cell resolution atherosclerotic (AS) information, there is a risk of systematic misjudgment in assessing the functional state of disease-associated immune cell subsets. That is, cell typing conclusions based on expression levels may not match the true functional state, leading to insufficient accuracy in downstream biomarker screening and target prediction.

[0023] Furthermore, taking tuberculosis (TB) as an example, existing analyses all use general-purpose AS databases (such as ASpedia and SpliceAid). These databases, however, contain shearing events constructed based on a broad tissue and disease context, without optimization for the specific immune microenvironment of Mycobacterium tuberculosis infection. Directly using these databases for TB-related sample analysis presents the following technical drawbacks: First, the coverage of TB-specific shearing events in general databases is low, leading to a large number of missed valid signals (high false negative rate); second, the general annotations contain numerous background noise events unrelated to TB, interfering with the screening of differential AS events and resulting in a high false positive rate.

[0024] Furthermore, existing mainstream single-cell transcriptome analysis workflows (such as Seurat and Scanpy systems) are designed with gene counting matrices as their core data structure. Their upstream alignment tools (such as STARsolo and Cell Ranger) do not include exon-level splicing quantification information by default. The read-level splicing site information required by AS quantification tools (such as rMATS and MAJIQ) is compressed and discarded in standard single-cell preprocessing workflows, resulting in inherent data incompatibility barriers at the technical level. That is, there is a structural gap in the input and output formats of existing toolchains, making it impossible for AS information to be captured by conventional workflows, resulting in the irreversible loss of key functional information.

[0025] To solve the above problems, refer to Figure 1This application provides a method for constructing a panel of biomarkers for single-cell asthma detection, comprising: S11: Perform single-cell transcriptome sequencing to obtain sequencing files, which record the geographical origin of each sample, including plains and plateau regions; This involves differentiating the geographical origin of patients when acquiring scRNA-seq data, as studies have shown that plateau and plain regions have different impacts on pathological mechanisms. To ensure that the final panel more accurately reflects this impact, this application labels and differentiates the geographical origin of samples during sequencing.

[0026] In this process, the sequencing data is integrated to record the batch origin, geographical origin (plateau / plain), and disease state (Active TB / Latent TB / HC) of each sample to obtain sequencing files, which serve as the basis for subsequent batch correction and stratification analysis.

[0027] S12: The sequencing file is processed to obtain target data, which includes: type label for each cell, cell barcode, and sorting BAM file for each sample; The process of processing the sequencing file to obtain the target data includes: S121: Calculate evaluation indicators based on the sequencing file, and remove data whose evaluation indicators are not within the preset requirements to obtain high-quality data. The evaluation indicators include Q30 base ratio, GC content distribution, and adapter contamination ratio. For example, data with a Q30 base ratio greater than or equal to the threshold are retained, and data with a Q30 base ratio less than the threshold are removed.

[0028] S122: For the high-quality data, intermediate data is obtained by genome comparison using a preset human genome as a reference genome. The intermediate data includes: a raw cell × gene count matrix and a sorted BAM file containing cell barcodes and UMI information. As a preferred implementation of this application, one tool is used to obtain the intermediate data, and then another tool is used for cross-validation. Consistent results are retained, and inconsistent results are removed.

[0029] S123: Obtain the quality control indicators of the intermediate data, and remove low-quality cells based on the quality control indicators. The quality control indicators include the number of genes detected per cell, the total UMI count per cell, and the mitochondrial gene ratio. The number of genes detected per cell is used to remove empty droplets (i.e., the number of genes detected per cell is less than the lower limit of the threshold range) and duplexes (i.e., the number of genes detected per cell is greater than the upper limit of the threshold range). Empty droplets do not contain real cells, and their "expression profile" is essentially background noise; if not removed, a large number of false cells will be introduced, contaminating subsequent clustering results. Cells with abnormally high gene counts are likely duplexes formed by two cells captured by the same droplet, and their mixed expression profiles will cause misclassification of cell types. The total UMI count per cell is used to remove low-quality cells (i.e., the total UMI count per cell is less than the corresponding threshold), and the mitochondrial gene ratio is used to remove apoptotic / damaged cells (i.e., the mitochondrial gene ratio is greater than the corresponding threshold).

[0030] S124: For data where low-quality cells have been removed, predictions are made, and cells predicted as doublets are removed. Prediction is performed using a pre-defined tool, such as DoubletFinder (v2.0.3). DoubletFinder is run independently for each sample to mark and remove cells predicted as doublets. This is because the above-mentioned filtering method based on quality control indicators can only remove obvious doublets with abnormally high gene counts, but cannot effectively identify "recessive doublets" with gene counts within the normal range. If such recessive doublets remain in the dataset, they will form a transitional pseudo-cell population between the two cell types in the dimensionality reduction clustering diagram, leading to errors in cell type annotation and serious biases in downstream differential gene analysis results.

[0031] S125: Batch integration and cell type annotation were performed on the data after removing diploid cells to obtain the target data. Batch integration and cell type annotation included batch effect correction, dimensionality reduction and clustering, and cell type annotation. Cell type annotation refers to the systematic collection and integration of cell subpopulation marker genes reported in published disease-related single-cell transcriptome studies to construct a tuberculosis-specific cell type annotation gene set, which is used to identify and label the biological identities of each cell population obtained from clustering.

[0032] S13: Input the sorted BAM file into various tools to obtain the PSI matrix; The process of inputting the sorted BAM file into various tools to obtain the PSI matrix includes: S131: Obtain the output of multiple tools separately; that is, this application uses multiple non-interoperable tools to jointly quantify in order to improve the coverage and reliability of AS event detection and overcome the problem of high false negative rate caused by insufficient sequencing depth in single-cell data by a single tool.

[0033] S132: Integrate the outputs of the multiple tools to obtain the original PSI matrix, where the rows of the original PSI matrix are AS events and the columns are cells; Furthermore, the process of integrating the outputs of the multiple tools to obtain the original PSI matrix includes: S1321: Convert the AS event coordinates in the output of the multiple tools into a standard format, wherein the AS event coordinates in the standard format include the AS event type; for example, convert the AS event coordinates output by multiple tools into a standard format: [AS Type]:[Chromosome]:[Start Position]-[End Position]:[Strand Direction] For example: SE:chr6:31234567-31245678:+ S1322: For AS events of type SE, A5SS, A3SS, or MXE, the event is retained only if a valid PSI value is detected in at least two preset tools; for AS events of type IR, the output of one preset tool is taken as the primary factor, and the event is retained after the output of the other preset tool is verified.

[0034] S133: The original PSI matrix that meets the following filtering conditions is retained to obtain the final PSI matrix, wherein the filtering conditions include: the cell coverage of each AS event is greater than or equal to the preset cell coverage, the PSI standard deviation of each AS event across all cells is greater than or equal to the preset standard deviation, and the number of valid AS events detected per cell is greater than or equal to the preset number of events.

[0035] Specifically, the cell coverage of each AS event is used to remove AS events with low detection rates or unreliable information. The PSI standard deviation of each AS event across all cells is used to remove AS events with no intercellular variation or no information content. The number of valid AS events detected per cell is used to remove low-quality cells with overly sparse AS information.

[0036] S14: Construct a variable splicing map based on the geographic origin, the cell type label, the cell barcode, and the PSI matrix; The construction of a variable splicing map based on the geographic origin, the cell type label, the cell barcode, and the PSI matrix includes: S141: Match and merge the PSI matrix with the cell type label and the geographic origin, using the cell barcode as the unique connection key, to construct a complete PSI annotation matrix containing geographic hierarchical information; S142: The mean cell type-specific PSI was calculated for each AS event, with cell type and geographic origin as the unit. S143: Within each cell type, compare the AS differences between plateau and plain regions and calculate the PSI difference; S144: Obtain a variable splicing map based on the complete PSI annotation matrix, the specific PSI mean, and the PSI difference.

[0037] In some alternative embodiments, the constructed variable splicing map can also be visualized.

[0038] S15: Perform differential AS marker screening and functional annotation on the variable splice map to obtain an AS detection marker panel.

[0039] First embodiment: The process of performing differential AS marker screening and functional annotation on the variable splice map to obtain an AS detection marker panel includes: S151: Within each cell type, calculate the PSI difference between plateau and plain regions and calculate the corrected P value; S152: Functional annotation is performed only on AS events where the absolute value of the PSI difference is greater than or equal to a preset PSI difference and the corrected P value is less than a preset corrected P value. The functional annotation includes: gene-level annotation, protein structure-level annotation, immune function-level annotation, and TB-specific association annotation. S153: For each candidate AS biomarker that has passed functional annotation, calculate a comprehensive priority score based on the importance of its biological function, and sort them in descending order of comprehensive priority score for each cell type, retaining the top K1 high-confidence AS biomarkers to obtain an AS detection biomarker panel, where K1 is a positive integer.

[0040] Second embodiment: Traditional methods for screening differential AS events (including statistical tests and classic machine learning methods such as LASSO regression and random forests) treat all AS events as independent features, ignoring the real regulatory dependencies between genes. If the network topology relationships between AS events are not considered, AS events that are not significant in individual statistical tests but play a crucial pivotal role in the regulatory network may be missed, resulting in insufficient biological integrity of the panel.

[0041] To overcome the aforementioned limitations, the second embodiment of this application provides an AS event feature selection method based on a Graph Neural Network (GNN). This method constructs a biological knowledge graph as a priori, naturally integrating a gene-gene interaction network into the information aggregation process. It utilizes the true dependencies between features to construct graph-structured data, and fuses node information through information propagation and aggregation operations. Its selection results outperform classical algorithms based solely on independent features in both feature quantity and classification accuracy. The specific implementation steps are as follows: Step 1: Construction of a Heterogeneous Biological Knowledge Graph A three-layer heterogeneous graph was constructed using AS events, host genes, and splicing factors (SF). The node set contained three types of nodes: AS event nodes, whose feature vectors were PSI matrices; host gene nodes, whose features were gene expression levels; and splicing factor nodes, whose features were their expression levels and known binding motif information. The edge set contained three types of edges: weighted co-expression edges based on gene-gene co-expression correlation coefficients; biological prior edges incorporating known pathway relationships from the GO / KEGG pathway database; and regulatory edges based on known regulatory relationships between splicing factors and their target AS events. The weights of each type of edge were normalized to prevent dimensional differences between edge types from affecting subsequent information aggregation.

[0042] This step integrates multi-omics data with biological priors represented by knowledge graphs, and uses GNNs to model the correlation structure between high-dimensional omics features, effectively reducing the data dimensionality and making it possible to analyze thousands of AS events simultaneously on hundreds of samples.

[0043] Step 2: Feature Learning for Graph Attention Network A Graph Attention Network (GAT) is employed to learn a neighbor-weighted aggregated representation for each AS event node. The attention weights reflect the contributions of different neighboring nodes to the importance of the current AS event, enabling the model to adaptively focus on biologically more relevant neighboring nodes, rather than treating all neighbors equally. After multi-layer graph convolutional aggregation, a low-dimensional embedding vector is generated for each AS event. This embedding vector simultaneously encodes the expressive features of the AS event itself and its topological location information within the regulatory network.

[0044] Step 3: Explainability Attribution After completing graph convolution feature learning, interpretable attribution methods (such as GNNExplainer and Integrated Gradients) are further integrated to quantitatively elucidate the contribution of each AS event node to the final classification task, and to extract key interactions between AS events and between AS events and splicing factors. This step is particularly crucial for identifying the regulatory network between splicing factors and AS events—without interpretable attribution, it is impossible to distinguish whether the importance of the features learned by the model stems from real biological regulatory relationships or statistical randomness in the data, thus affecting the biological credibility of the selected AS events.

[0045] Step 4: Feature Importance Extraction and Candidate Set Determination By combining the mean attention weight and gradient attribution score of each AS event node, the final feature importance comprehensive score is calculated. All AS events are sorted in descending order of the final feature importance comprehensive score, and the top-K AS events (K value is determined by cross-validation) are selected to enter the downstream panel construction process.

[0046] Third embodiment: AS event priority filtering based on Transformer self-attention mechanism Traditional single-priority scoring methods weight and combine multi-dimensional indicators into a single score. The weighting is subjective and fails to capture the long-range dependencies and nonlinear interactions between AS events. If the joint effects between AS events are ignored, a group of AS events that perform well individually but are highly redundant may be selected, resulting in low information utilization efficiency of the panel.

[0047] To overcome the above limitations, this invention provides an AS event prioritization method based on the Transformer self-attention mechanism. The Transformer, through its multi-head self-attention mechanism, can simultaneously capture the long-range dependencies and nonlinear interactions between AS events in a high-dimensional PSI matrix, making it naturally suitable for handling screening scenarios where complex joint effects exist between AS events. The specific implementation steps are as follows: Step 1: Input code Using the sample × AS event PSI matrix as input, each AS event is treated as an independent "token" in the sequence. A learnable positional encoding combined with AS event type encoding (six categories: exon skipping (ES), intron retention (RI), variable 5' splice site (AA), variable 3' splice site (AD), variable promoter (AP), and variable terminal exon (AT)) is used to construct the initial embedding vector for each AS event, enabling the model to distinguish the biological characteristics of different types of AS events.

[0048] Step 2: Multi-head self-attention feature interaction The initial embedding vectors are input into the Transformer encoder, and the attention weights between any two AS events are calculated using a multi-head self-attention mechanism to capture their joint dependency structure. To improve the training efficiency for long sequences, residual connections and layer normalization are used to stabilize the training process. The off-diagonal elements of the attention matrix reveal which combinations of AS events have the greatest joint contribution to clinical phenotype prediction, providing a basis for subsequent panel screening.

[0049] It should be noted that the core role of the Transformer model in this step is to capture the joint dependency structure between AS events and output attention weights, rather than directly using it for splice site sequence prediction. Existing research has shown that the Transformer-based deep learning framework outperforms the leading tool SpliceAI (PR-AUC = 0.820) on splice site detection tasks (PR-AUC = 0.834), verifying the effectiveness of the Transformer architecture in splice-related tasks and providing support for the feasibility of this approach.

[0050] Step 3: Attention Weight Attribution Screening The weight matrix of each attention head is extracted, and the comprehensive importance score of each AS event across all samples and all attention heads is calculated by cross-head mean aggregation and hierarchical attribution method. AS events that are significantly higher than the background level are selected into the candidate panel pool according to their importance scores.

[0051] Step 4: Task adaptation fine-tuning The Transformer model is fine-tuned on labeled samples using downstream supervised tasks (such as TB infection status classification and disease stage prediction) to enable the model to learn joint features of AS events directly related to the target clinical task. After fine-tuning, the fine-tuned attention weights are used as the final panel feature selection criteria to ensure that the selected AS events are highly correlated with the target diagnostic task.

[0052] It should be noted that the disease referred to in this application may be tuberculosis or other infectious diseases (such as COVID-19 or sepsis), and this application does not make any specific limitation.

[0053] To more clearly illustrate the proposed solution, a specific implementation method is provided below using tuberculosis as an example.

[0054] Tuberculosis (TB) is a chronic infectious disease caused by Mycobacterium tuberculosis (Mtb). There are approximately 10 million new cases worldwide each year, and it is one of the leading causes of death from a single infectious pathogen (WHO Global TB Report, 2023).

[0055] Tuberculosis (TB) in high-altitude areas exhibits significant epidemiological characteristics. In high-altitude regions of Tibet, Qinghai, and Yunnan in my country (≥2500m altitude), the incidence of TB is significantly higher than in plains areas. The hypoxic environment at high altitudes profoundly affects the metabolic state, functional phenotype, and transcriptional regulatory patterns of host immune cells by activating the hypoxia-inducible factor signaling pathway. Studies have shown that hypoxia conditions can directly affect the expression and activity of RNA splicing regulators, thereby altering the alternative splicing pattern of immune cells. Therefore, there may be fundamental differences in the immune response mechanisms between TB patients in high-altitude and low-altitude areas, and this difference has not yet been systematically studied at the level of alternative splicing.

[0056] Following Mtb infection, host immune cells (especially macrophages, T cells, and natural killer cells) undergo complex transcriptional reprogramming. In recent years, the widespread adoption of single-cell RNA sequencing (scRNA-seq) technology has enabled researchers to analyze the heterogeneity and dynamic changes of immune cells during TB infection at single-cell resolution, providing a technical basis for comparing immune differences between high-altitude TB and low-altitude TB.

[0057] Alternative splicing (AS) is an important mechanism for regulating gene expression in eukaryotes. Approximately 95% of human multi-exon genes undergo alternative splicing, resulting in protein isoforms with diverse functions. AS mainly includes the following types: 1. Exon Skipping (SE): The most common type of AS. 2. Intron Retention (IR) 3. Alternative 5' Splice Site (A5SS) 4. Alternative 3' Splice Site (A3SS) 5. Mutually Exclusive Exons (MXE) Studies have shown that atherosclerosis (AS) plays a crucial regulatory role in immune cell activation, differentiation, and inflammatory responses. Hypoxic environments can systematically alter cellular splicing patterns by regulating the expression of splicing factors such as SR proteins (Serine / Arginine-rich proteins) and hnRNP proteins. However, systematic research on the immune response to tuberculosis (TB) infection under high-altitude hypoxic conditions is extremely scarce, and a map of splicing differences between high-altitude and low-altitude TB remains elusive.

[0058] The closest existing technical solutions to this invention include: Option 1: Differential analysis of TB transcriptome based on bulk RNA-seq Representative study: Berry et al. (Nature, 2010) identified a set of interferon signaling gene signatures in the blood of TB patients using bulk RNA-seq.

[0059] Limitations: It cannot distinguish cell subpopulations and cannot analyze AS events.

[0060] Option 2: General single-cell transcriptome analysis workflow (such as Seurat, Scanpy) The existing workflow steps are: ① Raw reads quality control → ② Alignment with reference genome → ③ Cell clustering → ④ Differential gene expression analysis (DEG).

[0061] Limitations: The above process does not include an AS quantification module and cannot output information at the isoform level.

[0062] Option 3: Existing variable splicing analysis tools (rMATS, MAJIQ, VAST-TOOLS) The tools mentioned above can perform AS quantification on bulk RNA-seq data and output PSI values ​​(Percent Spliced ​​In).

[0063] Limitations: ① Designed specifically for bulk data, its application to single-cell data results in extremely high noise due to insufficient sequencing depth; ② Lacks TB-specific annotations; ③ Cannot be directly associated with cell type identity.

[0064] Option 4: Existing single-cell AS analysis tools (such as Psix, BRIE2) The tools mentioned above attempt to extract AS information from scRNA-seq data.

[0065] Limitations: ① Only a computational framework is provided, without any disease-specific annotations or references; ② Validation for TB immune cell subsets is not performed; ③ A panel of biomarkers that can be used directly is lacking.

[0066] On the one hand, most existing tuberculosis transcriptome studies only analyze immune cell status based on gene expression levels. However, alternative splicing events can independently regulate protein function and immune responses even when gene expression levels are not significantly different. In the absence of single-cell resolution AS information, there is a risk of systematic misjudgment in assessing the functional status of tuberculosis infection-related immune cell subsets. That is, cell typing conclusions based on expression levels may not match the true functional status, leading to insufficient accuracy in downstream biomarker screening and target prediction.

[0067] Secondly, the general-purpose AS database has poor specificity and large errors when used for tuberculosis analysis. Existing general-purpose AS databases (such as ASpedia and SpliceAid) contain cleavage events based on a broad tissue and disease context, without optimization for the specific immune microenvironment of Mycobacterium tuberculosis infection. Directly using these databases for TB-related sample analysis presents the following technical drawbacks: First, the coverage of TB-specific cleavage events in general databases is low, leading to a large number of missed valid signals (high false negative rate); second, the general annotations contain numerous background noise events unrelated to TB, interfering with the screening of differential AS events and resulting in a higher false positive rate.

[0068] Thirdly, existing single-cell workflows are technically difficult to integrate AS quantification, leading to information loss. Current mainstream single-cell transcriptome analysis workflows (such as Seurat and Scanpy systems) are designed with gene counting matrices as their core data structure. Their upstream alignment tools (such as STARsolo and Cell Ranger) do not include exon-level splicing quantification information by default. The read-level splicing site information required by AS quantification tools (such as rMATS and MAJIQ) is compressed and discarded in standard single-cell preprocessing workflows, resulting in inherent data incompatibility barriers at the technical level. That is, there is a structural gap in the input and output formats of existing toolchains, making it impossible for AS information to be captured by conventional workflows, resulting in irreversible loss of key functional information.

[0069] The specific proposal of this application is as follows: Module 1: Integration and Quality Control of Multi-Source TB Single-Cell Transcriptome Data 1.1 Sources and Integration of Original Data This invention integrates scRNA-seq data from tuberculosis patients from the following two sources: (1) Using the 3' end library construction protocol of the 10x Genomics Chromium platform, single-cell transcriptome sequencing was performed on TB patients in high-altitude areas and TB patients in plain areas.

[0070] Understandably, in addition to the 10x Genomics platform, data from platforms such as Smart-seq2 (full-length transcript sequencing with higher AS quantification accuracy) can be used to construct panels. Furthermore, besides PBMCs, data from lung tissue biopsy samples can be extended.

[0071] (2) During integration, record the batch origin, geographical origin (plateau / plain), and disease status (Active TB / Latent TB / HC) of each sample as the basis for subsequent batch correction and stratified analysis.

[0072] 1.2 Quality Control of Raw Reads Perform the following quality control steps on all self-tested raw FASTQ files (i.e., the sequencing files in 1.1): (1) Reads quality assessment ① Tool: FastQC (v0.11+) ②Evaluation indicators: Q30 base ratio (≥75%), GC content distribution, and connector contamination ratio. (2) Genome alignment ①Tool A: STARsolo (v2.7.10a) Reference genome: Human genome GRCh38 (hg38), annotation file: Ensembl v105 GTF Output: ① Original cell × gene count matrix; ② Sorted BAM file containing cell barcodes and UMI information.

[0073] ② Tool B: Cell Ranger (v7.0+, 10x Genomics official workflow) Reference genome: Human genome GRCh38 (hg38), annotation file: Ensembl v105 GTF Run in parallel with STARsolo to cross-validate the consistency of alignment results. (3) Single-cell level quality control filtration, filtering low-quality cells based on the three indicators in Table 1: Table 1 Quality control indicators Filtering threshold Filtering purpose Number of genes detected per cell (nFeature_RNA) 200-6000 Remove empty droplets (<200) and duplexes (>6000). Empty droplets contain no real cells, and their "expression profile" is essentially background noise; if not removed, a large number of false cells will be introduced, contaminating subsequent clustering results. Cells with abnormally high gene counts (>6000) are likely duplexes formed by two cells captured by the same droplet, and their mixed expression profiles can lead to misclassification of cell types. Total UMI count per cell (nCount_RNA) ≥500 Remove low-quality cells Mitochondrial gene ratio (percent.mt) <20% Remove apoptotic / damaged cells (4) Removal of twins Tool: DoubletFinder (v2.0.3) Operation: Run DoubletFinder independently for each sample, marking and removing cells predicted as doublets. Objective: The fixed threshold-based filtering method (step 3) can only remove obvious diatini with abnormally high gene counts, but cannot effectively identify "recessive diatini" with gene counts within the normal range. If such recessive diatini remain in the dataset, they will form a transitional pseudo-cell population between the two cell types in the dimensionality reduction clustering diagram, leading to cell type annotation errors and serious biases in downstream differential gene analysis results.

[0074] 1.3 Batch Integration and Cell Type Annotation (1) Batch effect correction (2) Dimensionality reduction and clustering (3) Cell type annotation: The system collects and integrates cell subpopulation marker genes reported in published tuberculosis-related single-cell transcriptome studies, constructs a tuberculosis-specific cell type annotation gene set, and uses it to identify and label the biological identity of each cell population obtained by clustering.

[0075] Final output of Module 1: High-quality cell × gene expression matrix Type label for each cell Sorted BAM file for each sample Module 2: Multi-tool combined for single-cell alternative splicing quantification and atlas construction 2.1 Quantification of single-cell AS events (multi-tool combined strategy) This invention employs a strategy of combining four complementary tools for quantification to improve the coverage and reliability of AS event detection, overcoming the high false negative rate caused by insufficient sequencing depth in single-cell data using a single tool.

[0076] Using the sorted BAM file generated by STARsolo in Module 1 as a unified input, run the following four tools respectively: Tool ①: rMATS (v4.1.2) Tool ②: MAJIQ (v2.4) Tool ③: IRFinder (v2.0) Tool 4: BRIE2 (v2.1) 2.2 Integration of PSI values ​​from multiple tools For the outputs of the four tools, the following integration strategy is implemented: (1) AS event ID unification Convert the AS event coordinates output by the four tools to a standard format: [AS Type]:[Chromosome]:[Start Position]-[End Position]:[Strand Direction] For example: SE:chr6:31234567-31245678:+ (2) Multi-tool consistency filtering For AS events of type SE, A5SS, A3SS, and MXE: a valid PSI value must be detected by at least two tools among rMATS, MAJIQ, and BRIE2 for the event to be retained. For AS events of the IR type: cross-validation is performed primarily using IRFinder results and secondarily using BRIE2 results.

[0077] Optionally, the AS event mining tool can be replaced with BRIE2, SCASL, etc. The PSI value calculation can be replaced with Splicing Ratio or Isoform Fraction (IF).

[0078] 2.3 PSI Matrix Quality Filtering The integrated raw PSI matrix (rows are AS events, columns are cells) was filtered according to the conditions shown in Table 2: Table 2 Filtering conditions threshold Filtering purpose Cell coverage per AS event ≥10% (meaning the event has a non-deleted PSI value in ≥10% of cells) Remove AS events with low detection rates and unreliable detection rates. Each AS event spans the PSI standard deviation across all cells. ≥0.1 Removal of AS events without intercellular variation and without information content Number of valid AS events detected per cell ≥50 Remove low-quality cells with excessively sparse AS information. 2.4: Construction of the TB Single-Cell Alternative Splicing Atlas (TB-scAS Atlas) (1) Integration of PSI matrix with cell type labels: The high-quality PSI matrix output in step 2.3 is matched and merged with the cell type label (Cell_Type) and geographic origin label (Geography: Highland / Lowland) output in module 1, and the cell barcode is used as the unique connection key to construct a complete PSI annotation matrix containing geographic stratification information.

[0079] (2) Calculation of mean PSI for cell type specificity Calculate for each AS event, using cell type and geographic origin as the unit. (3) Construction of the differential AS event matrix Within each cell type, the difference in AS between the high-altitude TB group and the plain TB group was compared, and delta PSI was calculated. (4) Visual presentation Module 3: Screening of Differential AS Markers and Multidimensional Functional Annotation 3.1 Filtering of Differential AS Events Within each cell type, the following statistical criteria were used to screen for significantly different AS events between high-altitude TB and low-altitude TB: |delta PSI| ≥ 0.1 and corrected P-value < 0.05. AS events that met all the above conditions proceeded to step 3-2 for functional annotation. 3.2 Hierarchical Functional Annotations For the differential AS events filtered in step 3-1, perform functional annotation at the following four levels in sequence, and integrate all annotation results into a unified annotation table: Level ①: Gene-level annotation Tools: Ensembl Biomart (R package bioomaRt v2.54) Information to obtain: Full gene name, Ensembl Gene ID, genomic coordinates (GRCh38), gene biological type (protein-coding / lncRNA, etc.) Pathway enrichment: GO enrichment analysis (goterm, BP / MF / CC ontologies) and KEGG pathway enrichment analysis (tool: clusterProfiler v4.6, FDR<0.05) were performed on the selected differentially expressed AS gene set. Level ②: Protein structure level annotation Tools: UniProt REST API + Pfam database (v35.0) Note: Determine whether the exons / introns affected by the AS event encode known protein functional domains, such as transmembrane domains, signal peptides, kinase domains, etc. Judgment criteria: There is ≥50% overlap between the genomic coordinates of the AS event and the Pfam functional domain annotation coordinates. Level ③: Annotation on Immune Function Level Database: ImmPort (https: / / www.immport.org) list of immune-related genes (Version: updated in 2023) Annotation content: Mark whether the gene to which the AS event belongs is a known immune-related gene, and annotate its immune function category (such as cytokines / chemokines, receptors, transcription factors, etc.). Level 4: TB-specific association annotation Database A: TBdb (Tuberculosis Database) – Retrieves host genes associated with Mtb infection. Database B: PubMed Literature Mining – Searching using "[gene name] AND tuberculosis AND splicing" as keywords to record existing literature-reported evidence of AS-TB association. Note: Indicate the strength of the known association between this gene and TB (direct literature evidence / indirect evidence / no known association). 3.3 Priority Scoring and Ranking of Markers For each candidate AS biomarker that passes functional annotation, a comprehensive priority score (PS) is calculated based on the importance of its biological function. The final retention criteria are as follows: for each cell type, the top 20–50 high-confidence AS biomarkers are retained in descending order of comprehensive priority score (the specific number is dynamically adjusted based on the total number of candidate biomarkers for each cell type to ensure that PS ≥ 6.0 / 10.0).

[0080] Module 4: Standardized Packaging and Verification of TB-scAS Marker Panel 4.1 Content Composition The core output of this invention—the TB plateau-scAS Marker—is stored in the form of a structured data table, with each record corresponding to a verified AS marker for the difference between plateau TB and plain TB.

[0081] 4.2 Independent clinical cohort validation To further verify the reliability of key AS biomarkers in the panel at the experimental level, this invention collected independent clinical samples and used reverse transcription quantitative PCR (RT-qPCR) to experimentally verify the AS biomarkers that ranked high in the priority score (PS) of the panel.

[0082] Based on the same inventive concept, this application provides a computer-readable storage medium storing a computer program or instructions thereon, wherein when the computer program or instructions are executed by a processor, the steps of the method for constructing a single-cell AS detection biomarker panel provided in any of the above embodiments are implemented.

[0083] Based on the same inventive concept, this application provides a panel construction device 20 for single-cell AS detection biomarkers, applying the panel construction method for single-cell AS detection biomarkers in the above embodiments. The device includes: The sequencing file acquisition module 21 is used to perform single-cell transcriptome sequencing to obtain sequencing files. The sequencing files record the geographical origin of each sample, including plains and plateaus. The target data acquisition module 22 is used to process the sequencing file to obtain target data, which includes: type label for each cell, cell barcode and sorting BAM file for each sample; PSI matrix acquisition module 23 is used to input the sorted BAM file into various tools to obtain the PSI matrix; AS map construction module 24 is used to construct a variable splice map based on the geographic origin, the cell type label, the cell barcode, and the PSI matrix; The marker panel construction module 25 is used to perform differential AS marker screening and functional annotation on the variable splice map to obtain an AS detection marker panel.

[0084] Based on the same inventive concept, this application provides a panel construction system for detecting biomarkers of single-cell AS disease, comprising: At least one processor and at least one memory; The memory stores the executable instructions of the processor; The processor is configured to execute the disease single-cell AS detection biomarker panel construction method provided in any of the above embodiments.

[0085] It is understood that the same or similar parts in the above embodiments can be referred to each other, and the contents not described in detail in some embodiments can be referred to the same or similar contents in other embodiments.

[0086] It should be noted that in the description of this application, the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance. Furthermore, in the description of this application, unless otherwise stated, "a plurality of" means at least two.

[0087] Any process or method described in the flowchart or otherwise herein can be understood as representing a module, segment, or portion of code comprising one or more executable instructions for implementing a particular logical function or process, and the scope of the preferred embodiments of this application includes additional implementations in which functions may be performed not in the order shown or discussed, including substantially simultaneously or in reverse order depending on the function involved, as will be understood by those skilled in the art to which embodiments of this application pertain.

[0088] It should be understood that various parts of this application can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using any one or a combination of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc.

[0089] Those skilled in the art will understand that all or part of the steps of the methods in the above embodiments can be implemented by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and when executed, the program includes one or a combination of the steps of the method embodiments.

[0090] Furthermore, the functional units in the various embodiments of this application can be integrated into a processing module, or each unit can exist physically separately, or two or more units can be integrated into a module. The integrated module can be implemented in hardware or as a software functional module. If the integrated module is implemented as a software functional module and sold or used as an independent product, it can also be stored in a computer-readable storage medium.

[0091] The storage media mentioned above can be read-only memory, disk, or optical disk, etc.

[0092] In the description of this specification, the references to terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0093] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for constructing a panel of biomarkers for single-cell AS detection, characterized in that, include: Single-cell transcriptome sequencing was performed to obtain sequencing files, which recorded the geographical origin of each sample, including plains and plateau regions; The sequencing file is processed to obtain target data, which includes: type label for each cell, cell barcode, and sorting BAM file for each sample; The sorted BAM file was input into various tools to obtain the PSI matrix; A variable splicing map is constructed based on the geographic origin, the cell type label, the cell barcode, and the PSI matrix; The variable splice map is subjected to differential AS marker screening and functional annotation to obtain an AS detection marker panel.

2. The method according to claim 1, characterized in that, The process of processing the sequencing file to obtain the target data includes: Evaluation metrics are calculated based on the sequencing files. Data that does not meet the preset requirements for evaluation metrics are removed to obtain high-quality data. The evaluation metrics include Q30 base ratio, GC content distribution, and adapter contamination ratio. For the high-quality data, intermediate data is obtained by comparing the genomes using a preset human genome as a reference genome. The intermediate data includes: a raw cell × gene count matrix and a sorted BAM file containing cell barcodes and UMI information. The quality control indicators of the intermediate data are obtained, and low-quality cells are removed based on the quality control indicators. The quality control indicators include the number of genes detected per cell, the total number of UMIs per cell, and the proportion of mitochondrial genes. The data for removing low-quality cells is used to predict and remove cells predicted to be dichotomous. The target data is obtained by batch integration and cell type annotation of the data after removing duplex cells. The batch integration and cell type annotation include batch effect correction, dimensionality reduction and clustering, and cell type annotation.

3. The method according to claim 1, characterized in that, The process of inputting the sorted BAM file into various tools to obtain the PSI matrix includes: Obtain the output of multiple tools separately; The outputs of the multiple tools are integrated to obtain the original PSI matrix, where rows represent AS events and columns represent cells; The original PSI matrix that meets the following filtering conditions is retained to obtain the final PSI matrix, wherein the filtering conditions include: the cell coverage of each AS event is greater than or equal to the preset cell coverage, the PSI standard deviation of each AS event across all cells is greater than or equal to the preset standard deviation, and the number of valid AS events detected per cell is greater than or equal to the preset number of events.

4. The method according to claim 3, characterized in that, The process of integrating the outputs of the multiple tools to obtain the original PSI matrix includes: The AS event coordinates in the output of the multi-tool are uniformly converted into a standard format, and the standard format includes the AS event type in the AS event coordinates. For AS events of type SE, A5SS, A3SS, or MXE, the event is retained only if a valid PSI value is detected in at least two preset tools; for AS events of type IR, the output of one preset tool is taken as the primary factor, and the event is retained after the output of the other preset tool is verified.

5. The method according to claim 1, characterized in that, The construction of a variable splicing map based on the geographic origin, the cell type label, the cell barcode, and the PSI matrix includes: The PSI matrix is ​​matched and merged with the cell type label and the geographic origin, and the cell barcode is used as the unique connection key to construct a complete PSI annotation matrix containing geographic hierarchical information. The mean cell type-specific PSI was calculated for each AS event, using cell type and geographic origin as the unit. Within each cell type, the differences in AS between plateau and plain regions were compared, and the PSI difference was calculated. A variable splicing map is obtained based on the complete PSI annotation matrix, the specific PSI mean, and the PSI difference.

6. The method according to claim 1, characterized in that, The process of performing differential AS marker screening and functional annotation on the variable splice map to obtain an AS detection marker panel includes: Within each cell type, the PSI difference between plateau and plain regions was calculated, and the corrected P value was also calculated. Functional annotation is performed only on AS events whose absolute PSI difference is greater than or equal to a preset PSI difference and whose corrected P value is less than a preset corrected P value. The functional annotation includes: gene-level annotation, protein structure-level annotation, immune function-level annotation, and TB-specific association annotation. For each candidate AS biomarker that passes functional annotation, a comprehensive priority score is calculated based on the importance of its biological function. For each cell type, the biomarkers are sorted in descending order of comprehensive priority score, and the top K1 high-confidence AS biomarkers are retained to obtain the AS detection biomarker panel, where K1 is a positive integer.

7. The method according to claim 1, characterized in that, The process of performing differential AS marker screening and functional annotation on the variable splice map to obtain an AS detection marker panel includes: A three-layer heterogeneous graph was constructed using AS events, host genes, and splicing factors. The node set contained three types of nodes: AS event nodes with PSI matrix as node feature vectors; host gene nodes with gene expression level as features; and splicing factor nodes with expression level and known binding motif information as features. The edge set contained three types of edges: weighted co-expression edges based on gene-gene co-expression correlation coefficients; biological prior edges based on known pathway relationships from the GO / KEGG pathway database; and regulatory edges based on known regulatory relationships between splicing factors and their target AS events. A graph attention network is used to learn a neighbor-weighted aggregation representation for each AS event node. The attention weights reflect the contributions of different neighboring nodes to the importance of the current AS event. After multi-layer graph convolution aggregation, a low-dimensional embedding vector is generated for each AS event. The embedding vector simultaneously encodes the expression features of the AS event itself and its topological location information in the control network. After completing graph convolution feature learning, interpretability attribution methods are further integrated to quantitatively clarify the contribution of each AS event node to the final classification task, and to extract key interaction relationships between AS events and between AS events and splicing factors. By combining the mean attention weight and gradient attribution score of each AS event node, the final feature importance comprehensive score is calculated. All AS events are sorted in descending order of the final feature importance comprehensive score, and the top K2 high-confidence AS events are retained to obtain the AS detection marker panel, where K2 is a positive integer.

8. The method according to claim 1, characterized in that, The process of performing differential AS marker screening and functional annotation on the variable splice map to obtain an AS detection marker panel includes: Using the PSI matrix of sample × AS event as input, each AS event is treated as an independent token in the sequence, and the initial embedding vector of each AS event is constructed by combining learnable position encoding with AS event type encoding. The initial embedding vector is input into the Transformer encoder, and the attention weights between any two AS events are calculated through a multi-head self-attention mechanism to capture their joint dependency structure. The weight matrix of each attention head is extracted. The comprehensive importance score of each AS event across all samples and all attention heads is calculated by cross-head mean aggregation and hierarchical attribution method. AS events that are significantly higher than the background level are selected into the candidate panel pool according to the importance score. The Transformer model is fine-tuned using downstream supervision tasks on labeled samples to obtain the fine-tuned attention weights. The candidate panel pool is filtered based on the attention weights to obtain the AS detection marker panel.

9. A panel construction device for single-cell AS disease detection biomarkers, characterized in that, include: The sequencing file acquisition module is used to perform single-cell transcriptome sequencing to obtain sequencing files. The sequencing files record the geographical origin of each sample, including plains and plateaus. The target data acquisition module is used to process the sequencing file to obtain target data, which includes: type label for each cell, cell barcode, and sorting BAM file for each sample; The PSI matrix acquisition module is used to input the sorted BAM file into various tools to obtain the PSI matrix. The AS map construction module is used to construct a variable splice map based on the geographic origin, the cell type label, the cell barcode, and the PSI matrix; The marker panel construction module is used to perform differential AS marker screening and functional annotation on the variable splice map to obtain an AS detection marker panel.

10. A panel construction system for single-cell AS disease detection biomarkers, characterized in that, include: At least one processor and at least one memory; The memory stores the executable instructions of the processor; The processor is configured to perform the method according to any one of claims 1-8.

Images