Single-cell transcriptome data-assisted AD analysis and classification method

Abstract

The invention relates to a single cell transcriptome data assisted AD analysis and classification method. The method comprises the following steps: acquiring a brain tissue sample from an AD patient and an age-matched healthy control, and treating the brain tissue by using a cell separation technology; then, sequencing the separated single cells by adopting a high-throughput single cell transcriptome sequencing platform, performing data quality control, and removing low-quality cells and low-expression genes; performing standardization processing on gene expression data, and eliminating technical deviation among different samples by using Log2 conversion and Z-score standardization; dimensionality reduction processing is carried out on the high-dimensional data by using dimensionality reduction algorithms of principal component analysis (PCA), t-SNE and UMAP; differential expression including DESeq2 and EdgeR is utilized to analyze and identify genes with significant difference between the AD patient and the healthy control; gSEA is adopted for gene set enrichment analysis, KEGG and GO are adopted for pathway analysis, and a key gene pathway and a biological process related to AD are determined; and inputting single cell transcriptome data of a new sample into the trained classification model, and performing AD classification prediction.

Description

Technical Field

[0001] The present invention relates to a method for single-cell transcriptome data-assisted AD analysis and classification. Background Art

[0002] Currently, methods for analyzing and classifying Alzheimer's disease (AD) are constantly evolving, aiming to improve the early diagnosis and classification accuracy of the disease through different data sources and analysis techniques. However, existing methods, such as the AD classification method based on multi-instance learning and multi-scale feature fusion described in Chinese invention patent CN118196538A, although they have improved the accuracy and generalization ability of classification to a certain extent, still have some significant shortcomings.

[0003] First of all, Chinese patent CN118196538A is mainly based on structural magnetic resonance imaging (sMRI) data. Although the application of this imaging method in AD classification is mature, it has some inherent limitations. sMRI mainly provides brain structural information. Although changes in certain areas of the brain can be captured through plaque segmentation and multi-scale feature extraction, sMRI data cannot directly reflect pathological changes at the molecular level. This means that sMRI can only capture relatively late brain structural changes, but cannot reveal biological markers in the early stages of the disease. Therefore, although this method can identify brain areas that contribute significantly to the classification results, changes in these areas are often the result of lesions rather than the beginning of lesions. Therefore, this method is insufficient in sensitivity for early diagnosis of the disease and cannot effectively identify early molecular signals of AD development.

[0004] Secondly, the analysis method based on sMRI is highly dependent on image processing and feature extraction technology, which may introduce additional noise and errors during the processing. For example, in the method of Chinese patent CN118196538A, sMRI images need to undergo multiple plaque segmentation and feature fusion, and these steps depend on the accuracy of the algorithm and the quality of the data. In actual operation, the limitations of sMRI image resolution, the variability in the image acquisition process, and individual differences among patients may lead to inaccuracy in feature extraction. This method that relies on multi-step processing is prone to introduce cumulative errors, thereby affecting the final classification results. In addition, the selection of large and small plaque models and feature fusion strategies involved in the plaque segmentation process are also subjective and complex, which may lead to inconsistent performance in different data sets or different populations, reducing the generalization ability of the method. Summary of the invention

[0005] The purpose of the present invention is to provide a method for single-cell transcriptome data-assisted AD analysis and classification, thereby solving some of the drawbacks and shortcomings pointed out in the background technology.

[0006] The present invention solves the above-mentioned technical problems by adopting the following technical solution: a method for single-cell transcriptome data-assisted AD analysis and classification, comprising:

[0007] S1. Sample collection and preparation:

[0008] S1.1. Obtain brain tissue samples from AD patients and age-matched healthy controls; use a brain tissue biobank or obtain samples through a brain tissue donation program;

[0009] S1.2. Use cell separation techniques including flow cytometry and FACS to process brain tissue and obtain single cell suspension; separate different types of brain cells including neurons, astrocytes, and microglia through mechanical and enzymatic methods;

[0010] S2. Single-cell transcriptome sequencing:

[0011] S2.1. Sequence the isolated single cells using 10xGenomicsChromium's high-throughput single-cell transcriptome sequencing platform to obtain the gene expression profile of each cell.

[0012] S2.2. Perform data quality control to remove low-quality cells and low-expression genes; use CellRanger's data analysis tool to process the raw sequencing data and generate a gene expression matrix;

[0013] S3. Data preprocessing:

[0014] S3.1. Standardize the gene expression data, including using Log2 transformation and Z-score normalization, to eliminate technical bias between different samples;

[0015] S3.2. Use dimensionality reduction algorithms including principal component analysis (PCA), t-SNE, and UMAP to reduce the dimensionality of high-dimensional data;

[0016] S4. Feature extraction:

[0017] S4.1. Differential expression analysis including DESeq2 and EdgeR was used to identify genes with significant differences between AD patients and healthy controls.

[0018] S4.2. Gene set enrichment analysis was performed using GSEA and pathway analysis including KEGG and GO to identify key gene pathways and biological processes related to AD;

[0019] S5. Classification and diagnosis of AD:

[0020] S5.1. Input the single-cell transcriptome data of the new sample into the trained classification model to perform AD classification prediction;

[0021] S5.2. Output the classification results and combine them with the patient's clinical data and other biomarkers including Aβ and tau protein levels to provide comprehensive auxiliary diagnosis recommendations.

[0022] Furthermore, the data quality control method includes:

[0023] S1. Removal of low-quality cells:

[0024] Calculate the total UMI, UniqueMolecularIdentifier count U and number of detected genes G for each cell; using the formula:

[0025]

[0026] Calculate cell quality score Q c , cells below the preset threshold are considered low-quality cells;

[0027] Remove cells with too high mitochondrial gene expression ratio MGR; use the formula:

[0028]

[0029] Among them, M mito is the total UMI number of mitochondrial gene expression, M total is the total UMI number of all genes expressed in the cell; set a threshold θ, and remove the cell when MGR>θ;

[0030] Low-quality cells that do not meet the standards are removed using cell filtering criteria including UMI counts, gene numbers, and mitochondrial gene ratios; the comprehensive score S is calculated as follows:

[0031]

[0032] Among them, α and β are weight parameters; cells below the set threshold are removed;

[0033] S2. Removal of low-expression genes:

[0034] Calculate the expression frequency F of each gene in all cells; use the formula:

[0035]

[0036] Among them, F g is the expression frequency of gene g, E gi is the expression level of gene g in the i-th cell, N is the total number of cells; set the gene expression threshold δ to remove genes expressed in less than the percentage of δ cells; the removal condition is: if F g <δ, then remove gene g.

[0037] Furthermore, the raw sequencing data FASTQ file is processed using CellRanger software to generate a gene expression matrix;

[0038] The specific steps are as follows:

[0039] S1. Read the original sequencing data and perform gene sequence alignment; CellRanger uses splicing alignment to align reads to the reference genome using the formula:

[0040]

[0041] Among them, A i is the alignment score of the i-th read, S i is the number of successfully aligned gene fragments, R i is the total number of reads;

[0042] S2. Then, the UMI tags are deduplicated to generate redundant gene expression counts. The UMI deduplication formula is:

[0043]

[0044] Among them, U unique is the UMI count after redundancy removal, M is the number of all UMI tags, and N j is the repeat count of the jth UMI label;

[0045] S3, generate the initial gene expression matrix, and record the expression level of each gene in each cell;

[0046] The expression matrix generation formula is:

[0047]

[0048] Among them, E gi is the expression level of the g-th gene in the i-th cell, C gi is the UMI count of the g-th gene in the i-th cell, U total is the total UMI count in the i-th cell.

[0049] Furthermore, the method for eliminating the technical deviation between different samples includes:

[0050] S1. First, based on the Poisson distribution characteristics of the UMI counts of single-cell data, an improved Log transformation is performed on the UMI counts in each cell using the equation:

[0051]

[0052] Among them, L giis the improved Log-transformed expression value of the g-th gene in the i-th cell, C gi is the original UMI count; by smoothing small UMI count values, the impact of noise is reduced;

[0053] S2. Then before Z-score normalization, a weight factor is introduced based on the frequency of each gene in all cells, using the equation:

[0054]

[0055] Among them, W g is the weight of gene g, F g is the frequency of gene g in all cells, k and τ are parameters that adjust the shape of the curve;

[0056] S2.1. Data standardization using weighted Z-score standardization:

[0057]

[0058] Among them, Z gi is the normalized expression value of the g-th gene in the i-th cell; μ g is the weighted average Log2 expression value of gene g in all cells; σ g is the standard deviation based on the weights; N is the total number of cells;

[0059] Among them, μ g and σ g are the weighted mean and standard deviation, respectively, and are calculated as follows:

[0060]

[0061] The contribution of each gene was adjusted by weighting the normalized gene expression frequency.

[0062] Furthermore, the method for identifying genes with significant differences between AD patients and healthy controls was constructed as follows:

[0063] S1. Use a pseudo-counting method based on Gamma distribution adjustment; adjust each count according to the potential distribution characteristics of gene expression, and improve the pseudo-counting formula:

[0064]

[0065] Where: γ is the Gamma function used to generate the adjustment coefficient; λ and k are shape and scale parameters dynamically calculated based on the global characteristics of gene expression data;

[0066] S2. Due to batch effects and differences in cell size in single-cell data, a composite normalization method based on a multivariate regression model was used to adjust for multiple deviations from experimental design and sample processing; the composite normalization formula was used:

[0067]

[0068] Among them: β0, β1, β2 are coefficients estimated by regression analysis, representing the effects of baseline correction, cell-specific factors, and batch effects, respectively;

[0069] S3. When using EdgeR for differential expression analysis, the estimation of dispersion parameters is optimized by introducing dynamic local regression technology. The dynamic local regression dispersion estimation formula is:

[0070] φ adjusted =LOESS(φ g ,{x i},{y i},θ)

[0071] Where: LOESS represents the local weighted scatter point smoothing function; {x i},{y i} represent the gene expression and observed dispersion data points, respectively; θ is a dynamically selected smoothing parameter that is adjusted based on the local density and variability of the data.

[0072] Furthermore, the construction of a model for single-cell transcriptome data-assisted AD analysis and classification includes the following steps:

[0073] S1. Select machine learning algorithms including random forest, support vector machine, and neural network to build a classification model;

[0074] S2, using the extracted differential gene expression data as feature input for model training; using K-fold cross validation to optimize model parameters;

[0075] S3. Verify the model performance through an independent test set and evaluate the accuracy, sensitivity, specificity and stability of the model; use indicators including AUC-ROC curve and confusion matrix to evaluate the model.

[0076] Furthermore, the model training and model parameter optimization method includes:

[0077] S1. First, an improved differential expression analysis method was used, using feature selection based on negative binomial distribution, where the significance evaluation formula is:

[0078]

[0079] Among them, S giis the significance score of the g-th gene in the i-th cell, C gi is the raw count, k and θ are the parameters of the negative binomial distribution;

[0080] S1.1. Simultaneously combine nonlinear dimensionality reduction techniques including autoencoders to capture complex patterns in gene expression data. The loss function of the autoencoder is:

[0081]

[0082] Among them, L AE is the loss function of the autoencoder, x i is the input data, f θ is the network function, θ(t) is the time-dependent regularization parameter, and λ is the regularization coefficient;

[0083] S2. A deep learning model is used and its structure is adjusted to adapt to the single-cell data characteristics including sparsity and intercellular heterogeneity. Model optimization is achieved through dropout technology, and the dropout ratio is controlled by the following function:

[0084]

[0085] Where D(x) is the output after dropout, α and β are control parameters;

[0086] S2.1. Use a customized loss function to solve the problem of class imbalance. The loss function is:

[0087]

[0088] Among them, L class is the classification loss function, w c is the class weight, y c and are the true and predicted labels respectively, μ is the smoothing regularization coefficient;

[0089] S3. Finally, the K-fold validation strategy based on clustering is implemented to ensure the representativeness and generalization ability of each compromise. The clustering division is performed by the following formula:

[0090]

[0091] Among them, K(x i ,x j ) is the similarity score between samples i and j, γ and η are adjustment parameters;

[0092] S3.1. Optimize model parameters by combining multiple evaluation indicators at the same time. The comprehensive evaluation function is:

[0093]

[0094] Among them, Q is the comprehensive evaluation function, ω m is the weight of the mth evaluation index, φ m (t) and ψ m (t) are the time correlation functions of the mth evaluation index.

[0095] Beneficial effects of the present invention:

[0096] 1. Improved accuracy: By combining a variety of advanced data processing technologies and deep learning models, the present invention can identify significantly different genes between AD patients and healthy controls at the single-cell level, significantly improving the accuracy of the classification model. This enables the model to more accurately reflect the biological characteristics of the disease when predicting and diagnosing AD.

[0097] 2. Effective control of data noise: The present invention effectively eliminates technical deviations, batch effects and cell size differences in single-cell transcriptome data through improved pseudo-count adjustment, composite normalization and dynamic local regression technology, reduces the impact of data noise, and thus improves the reliability of analysis results.

[0098] 3. Strong generalization ability: Through the clustering-based K-fold validation strategy, the present invention can ensure that each compromised sample is representative, thereby enhancing the generalization ability of the model. Even under different data sets or different sample types, the model can still maintain stable classification performance.

[0099] 4. Highly adaptable model structure: By customizing the structure and loss function of the deep learning model, the present invention can handle the sparsity and heterogeneity of single-cell data, while addressing the problem of class imbalance, ensuring that the model performs well in various complex data environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0100] Figure 1 This is a flow chart of the method for single-cell transcriptome data-assisted AD analysis and classification of the present invention.

[0101] Figure 2 The CellRanger software was used to process the original sequencing data FASTQ file and generate a gene expression matrix step flow chart.

[0102] Figure 3 A flow chart was constructed for the method of the present invention to identify genes that are significantly different between AD patients and healthy controls.

[0103] Figure 4 This is a flow chart of the model training and model parameter optimization method of the present invention. DETAILED DESCRIPTION

[0104] The specific implementation modes of the present invention will be described in detail below in conjunction with the accompanying drawings.

[0105] As attached Figure 1 As shown, the first step of the present invention involves obtaining brain tissue samples from Alzheimer's disease (AD) patients and age-matched healthy controls. The samples can be obtained through a brain tissue biobank or collected through a brain tissue donation project to ensure that there is enough sample volume to represent the diversity of the disease and its distribution in the population; next, the brain tissue samples are processed using cell separation techniques such as flow cytometry (FACS) to obtain a single cell suspension, including separating different types of brain cells from the samples by mechanical and enzymatic methods, such as neurons, astrocytes, microglia, etc. The successful separation of these cell types is crucial for subsequent single-cell transcriptome analysis. They each have a unique contribution to AD, and understanding how these cell types change during the course of the disease is the key to understanding AD.

[0106] The second step is single-cell transcriptome sequencing. The 10x Genomics Chromium high-throughput single-cell transcriptome sequencing platform is used to sequence single cells isolated from AD patients and healthy controls. This process can obtain the gene expression profile of each cell, and researchers can observe and analyze differences in gene expression at extremely high resolution. This information is crucial for understanding the molecular mechanism of AD. Data quality control is then performed, including removing cells and low-expression genes that show low quality due to sample preparation, sequencing or other technical problems, to ensure that only reliable information is contained in the data set. In addition, the use of CellRanger software to process raw sequencing data can effectively correspond sequencing read information to the reference genome and generate a detailed gene expression matrix, which lays a solid foundation for subsequent bioinformatics analysis and statistical processing.

[0107] The third step covers data preprocessing. First, the gene expression data are standardized using Log2 transformation and Z-score normalization. Log2 transformation helps scale the wide range of gene expression data to a more manageable and standardized scale, which can reduce the bias of a single highly expressed gene on the analysis results. Z-score normalization further processes these data by adjusting the mean of each gene expression to 0 and the standard deviation to 1, thereby eliminating technical bias between different samples, which is especially important for subsequent statistical tests and comparative analyses. Dimensionality reduction algorithms such as principal component analysis (PCA), t-SNE, and UMAP are then used to process high-dimensional data. These algorithms can effectively extract the main sources of variation from complex data in order to better visualize and further analyze the data structure. PCA focuses on capturing the components with the largest variance in the data, while t-SNE and UMAP are better at revealing population structures and nonlinear relationships in the data, which is particularly critical for understanding cell-to-cell heterogeneity and identifying potential biological patterns.

[0108] The fourth step is feature extraction to accurately identify and analyze AD-related biomarkers. First, two differential expression analysis tools, DESeq2 and EdgeR, are used to identify significantly different genes between Alzheimer's disease (AD) patients and healthy controls. These tools are based on statistical models, such as negative binomial distribution, to evaluate the difference in expression levels of each gene between disease and control samples, allowing researchers to effectively screen candidate genes related to the disease process from thousands of genes. Gene set enrichment analysis (GSEA) and pathway analysis tools, such as KEGG and GO, were then used to further analyze the biological functions of these differentially expressed genes and the metabolic or signal transduction pathways involved. The GSEA method statistically evaluates whether a predefined gene set (usually based on functional or pathway annotations) is enriched at the top or bottom of the gene expression ranking list, thereby revealing the biological processes or pathways most relevant to the disease state. KEGG and GO analysis provides a framework to systematically understand how these genes work together to affect specific biological processes or cellular functions. Through these complex bioinformatics analysis methods, researchers can map out key gene pathways and biological processes related to AD, providing a scientific basis for further mechanism research and the discovery of potential therapeutic targets.

[0109] The fifth step is to use the trained machine learning model to classify and diagnose AD: first, the single-cell transcriptome data from the new sample is input into the model to make AD classification predictions. This process involves using the model to interpret and analyze the gene expression pattern of each cell to determine whether the cell belongs to an AD patient or a healthy control; next, the classification results output by the model are further analyzed, including combining the patient's clinical data such as symptoms, disease progression information, and information on other key biomarkers (such as Aβ and tau protein levels). The combination of these data can verify the accuracy of the prediction and provide more targeted and personalized auxiliary diagnostic recommendations. The application of this method helps to improve the accuracy of AD diagnosis and provide patients with more effective treatment guidance and management.

[0110] The model construction process of single-cell transcriptome data-assisted AD analysis and classification includes: first, selecting suitable machine learning algorithms to build classification models, including random forests, support vector machines, and neural networks. Each algorithm has its own unique advantages. For example, random forests perform well in processing high-dimensional data and preventing overfitting, while support vector machines are good at processing complex nonlinear decision boundaries, and neural networks are particularly suitable for extracting deep patterns and features from large amounts of data; next, the gene expression data extracted from differential expression analysis are input into the model as features for training. These feature data can capture the significant genetic differences between AD patients and healthy controls. K-fold cross-validation is used to optimize model parameters during model training. The generalization ability of the model is ensured by repeatedly dividing the training set and the validation set, and the optimal parameter combination is found; finally, the performance of the model is verified by an independent test set, which can objectively evaluate the accuracy, sensitivity, specificity, and stability of the model in practical applications. Commonly used evaluation indicators include AUC-ROC curves and confusion matrices. The AUC-ROC curve can intuitively show the performance of the model at different decision thresholds, while the confusion matrix provides a more detailed analysis of the classification results to help identify potential weaknesses and room for improvement of the model.

[0111] Embodiment 1:

[0112] The pathogenesis of Alzheimer's disease (AD) is being studied, and 100,000 single cells were extracted from brain tissues of 10 AD patients for transcriptome sequencing. First, data quality control is required to ensure that only high-quality cell data are analyzed. During the data analysis process, it was found that the UMI count of the cell (that is, the number of Unique Molecular Identifiers per cell) and the number of genes detected are two important indicators for evaluating cell quality.

[0113] Removal of low-quality cells

[0114] First, calculate the total UMI count U and the number of detected genes G for each cell. In this case, set the U of a cell to 5000 and G to 3000. According to the formula:

[0115]

[0116] The quality score Q of the cell can be calculated c .

[0117] Substituting U=5000 and G=3000 into the formula:

[0118]

[0119] If the preset quality threshold is 4000, then the quality score of this cell is higher than the threshold and is therefore considered a high-quality cell.

[0120] Next, calculate the mitochondrial gene expression ratio MGR. Set the total number of UMIs expressed by mitochondrial genes in the cell M mtio is 300, the total number of UMIs expressed by all genes is M total is 5000, then the calculation formula of MGR is:

[0121]

[0122] The threshold θ is set to 5%. Since the MGR of this cell is 6%, which is higher than the threshold, this cell is considered to be of poor quality and should be removed. However, the comprehensive score S also needs to be considered to further decide whether to retain the cell.

[0123] The calculation formula of the comprehensive score S is:

[0124]

[0125] Assuming the weight parameters α is 0.7 and β is 0.3, the comprehensive score S is calculated as follows:

[0126]

[0127] The comprehensive score threshold is set to 3000, and since S=2886.452 is lower than the threshold, the cell is removed.

[0128] It can be seen that this data quality control method comprehensively evaluates the quality of each cell by calculating indicators such as UMI count, gene number, and mitochondrial gene expression ratio, combined with weight parameters. Setting a reasonable threshold can effectively remove low-quality cells and ensure the reliability and accuracy of subsequent analysis.

[0129] To further improve the quality of the data and ensure that the model training is based only on valid and meaningful gene expression data, it is necessary to remove genes that are not significantly expressed in the vast majority of cells. In previous experiments, 100,000 single cells were extracted from the brain tissues of 10 AD patients, and the expression of 20,000 different genes was measured in these cells.

[0130] Removal of low-expressed genes

[0131] First, calculate the expression frequency F of each gene in all cells g . Assume that a gene g is expressed in these 100,000 cells and its expression level in each cell is expressed by E gi The expression frequency of the gene is calculated by the following formula:

[0132]

[0133] Where N is the total number of cells, i.e. 100,000. Set the expression level of gene g in these cells to E g1 ,E g2 ,...,E gN They are 5, 3, 0, 7, 4, etc., and the sum is 5000. Then F g for:

[0134]

[0135] This result shows that the expression frequency of gene g in total cells is 0.05, which means that the gene is expressed in 5% of cells. If the gene expression threshold δ is set to 10%, then due to F g If the value is lower than δ, the gene g will be considered as a low-expression gene and needs to be removed in subsequent analysis.

[0136] For further explanation, let δ = 5%, and the expression frequency F of gene h h If the expression of p-value is 0.07 (i.e. 7%), this gene will be above the threshold and will therefore be retained in the subsequent analysis. This gene removal strategy ensures that the genes of interest are expressed in a relatively large proportion of cells, thereby improving the effectiveness of the analysis.

[0137] In this way, genes that are not significantly expressed in the vast majority of cells are effectively removed, thereby reducing data noise and improving the efficiency and accuracy of model training. Setting a reasonable δ value (e.g., between 5% and 10%) is crucial to balancing the removal of meaningless data and retaining sufficient information.

[0138] Embodiment 2:

[0139] Combination Figure 2 In this embodiment, the original sequencing data is processed using CellRanger software to generate a gene expression matrix, including starting from the original FASTQ file and converting the sequencing data into a gene expression matrix that can be used for downstream analysis through a series of algorithms.

[0140] Single-cell transcriptome data were extracted from brain tissue samples of AD patients, generating a large number of FASTQ files. These files record millions of short sequence fragments (reads), each of which represents a gene expression fragment. In order to align these reads with the reference genome, the splicing alignment function in the CellRanger software is required.

[0141] Read the raw sequencing data and perform gene sequence alignment. In this process, each read will be compared with the gene sequence of the reference genome to determine its source. Set the reads of a sample to be aligned and obtain the following data:

[0142] Total number of reads R i is 100,000;

[0143] The number of gene fragments successfully aligned to the reference genome S i is 85,000.

[0144] Using the formula:

[0145]

[0146] The alignment score A for each read can be calculated i :

[0147]

[0148] This result shows that in this sample, 85% of the reads were successfully aligned to the reference genome, indicating that the sequencing quality of this sample is high and the data can be further used for subsequent analysis.

[0149] Next, a similar comparison was performed on another sample, and the following results were obtained:

[0150] Total number of reads R i is 100,000;

[0151] The number of successfully aligned gene fragments S i is 60,000.

[0152] The alignment score A is calculated using the same formula i :

[0153]

[0154] This low alignment score indicates that there are issues with the sequencing quality of this sample and further inspection or filtering is required.

[0155] In the actual analysis process, a threshold of the alignment score is usually set, such as 0.7, to screen out high-quality samples. If the alignment score of a sample is lower than this threshold, it will be marked as low quality and require resequencing or data processing.

[0156] It can be seen that the splicing alignment function in the CellRanger software can effectively evaluate the quality of sequencing data. The alignment score A involved in the formula iBeing able to quantify the alignment quality of each sample provides an important basis for subsequent data processing and analysis. In practical applications, this method has been proven to be feasible and effective, especially when processing large-scale single-cell sequencing data, it can significantly improve the accuracy of the data and the reliability of downstream analysis. Setting an appropriate alignment score threshold (e.g., a range between 0.7 and 0.85) is key to ensuring data quality.

[0157] Next, duplicates are removed based on the UMI (Unique Molecular Identifier) ​​tags to generate redundant gene expression counts.

[0158] In single-cell transcriptome sequencing, UMI tags are used to identify and remove duplicate reads during PCR amplification to ensure the accuracy of gene expression.

[0159] De-duplication of UMI labels is performed to generate redundant gene expression counts. Assume that when processing a specific sample, 3 UMI labels are detected, and the duplicate counts of each UMI label are N1=5, N2=3, and N3=2. The following formula is used to calculate the UMI count after de-duplication:

[0160]

[0161] Among them, U unique is the UMI count after de-redundancy, M is the total number of all UMI labels, and N j is the repeat count of the j-th UMI label.

[0162] Substitute the above data to calculate:

[0163]

[0164] First, calculate the value of each part:

[0165]

[0166] Add these values ​​together to get the de-redundant UMI count:

[0167] U unique =0.2+0.3333+0.5=1.0333

[0168] The results show that in this sample, the UMI count after de-redundancy is 1.0333. This value indicates that during the processing, taking into account the repeatability of PCR amplification, the actual independent UMI count is approximately 1.0333.

[0169] To ensure the reliability of the results, it is necessary to set an appropriate UMI deduplication threshold. In practical applications, the lower threshold of the UMI count is usually set between 1.0 and 1.5 to remove low-quality data with severe repeat amplification. Through this process, more accurate gene expression counts can be obtained, further improving the reliability and accuracy of downstream analysis.

[0170] The final step is to generate the initial gene expression matrix, which is the basis for all subsequent analyses and records the expression level of each gene in each cell.

[0171] Generate an initial gene expression matrix.

[0172] Assume that 1000 single cells have been extracted from a sample of an AD patient, and the preliminary data quality control and UMI deduplication have been completed. Now we need to generate an expression matrix to record the expression level of each gene in each cell. Assume that in a certain cell, the UMI count of the gth gene is C gi =150, and the total UMI count in the cell is U total =5000.

[0173] The expression level of the g-th gene in the i-th cell is calculated using the following formula:

[0174]

[0175] Among them, E gi is the expression level of the g-th gene in the i-th cell, C gi is the UMI count of the g-th gene in the i-th cell, U total is the total UMI count in the i-th cell.

[0176] Substitute the data for calculation:

[0177]

[0178] The results show that in this cell, the expression level of gene g is 30,000. This value represents the expression level of this gene in this specific cell after normalization. This expression calculation method allows direct comparison between different cells, and it has been normalized to eliminate the bias introduced by differences in total UMI counts of cells.

[0179] Let's continue with the example of 1,000 cells in the entire sample. We assume that similar calculations are performed in each cell, and an expression matrix containing all cells and genes is obtained. This matrix can be used for subsequent analysis steps, including feature extraction, training of classification models, and identification of AD-related genes.

[0180] The value range of the coefficient, for example, U total, typically ranging from thousands to millions, depending on the sequencing depth of each cell and the quality of the biological sample.

[0181] Embodiment 3:

[0182] In this embodiment, the UMI count is adjusted to reduce the impact of noise and bias on the analysis results, thereby improving the accuracy of the AD classification model.

[0183] Suppose that a single-cell transcriptome dataset containing samples from AD patients and healthy controls is being analyzed. For a specific gene g, the UMI count C is obtained from the cells. gi = 20. To eliminate technical bias, the counts will be processed using a modified Log transform.

[0184] First, based on the Poisson distribution characteristics of the UMI count of single-cell data, an improved Log transformation is performed on the UMI count in each cell. The specific equation is as follows:

[0185]

[0186] Among them, L gi is the improved Log-transformed expression value of the g-th gene in the i-th cell, C gi Count of raw UMIs.

[0187] C gi =20Substitute into the formula to calculate:

[0188]

[0189] L gi =log212.4495

[0190] L gi ≈3.635

[0191] The improved Log-transformed expression value of gene g in this cell was calculated to be 3.635. This value was smoothed to reduce the noise caused by small UMI counts, making the expression of this gene more consistent and comparable in cross-cell analysis.

[0192] There are also multiple cells in the data set where the UMI counts of the gene are 15, 5, and 0. Similarly, these values ​​are substituted into the above formula for calculation to obtain the corresponding Log-transformed expression values ​​and generate a standardized expression matrix.

[0193] These normalized expression values ​​can be used as input data for subsequent analyses. By reducing technical bias between different samples, AD classification models will be able to more accurately identify disease-associated gene expression patterns.

[0194] The specific values ​​of the coefficients and thresholds will be determined based on the characteristics of the actual dataset. For example, in this calculation, the Poisson distribution characteristics and technical bias determine how to select the raw UMI count C gi and improved Log transformation parameters. In practical applications, the selection of these coefficients needs to be based on statistical analysis of the data set and experimental design, and the values ​​usually range from a few to dozens, depending on the sequencing depth and the characteristics of the biological sample.

[0195] The data is then Z-score normalized, and a weighting factor based on the frequency of gene occurrence in all cells is introduced in the process. Assume that the expression values ​​of gene g in several cells have been obtained by modified Log transformation in the previous step. Now, these expression values ​​are used for weighted Z-score normalization.

[0196] First, introduce the weight factor W of gene g g . Weight factor W g The calculation formula is as follows:

[0197]

[0198] Among them, W g is the weight of gene g, F g is the frequency of occurrence of gene g in all cells, and k and τ are parameters that adjust the shape of the curve. In practical applications, the values ​​of parameters k and τ usually range from 1 to 10, and the specific values ​​depend on the extensiveness of gene expression and the distribution of cell samples.

[0199] Set the frequency of occurrence of gene g to F g is 0.6, k is 2, and τ is 0.5, then the weight factor W of gene g is g The calculation is as follows:

[0200]

[0201] W g ≈0.55

[0202] By calculation, we get the weight factor W of gene g g Next, the weighted Z-score normalization formula will be used to normalize the expression value of gene g.

[0203] The weighted Z-score normalization formula is as follows:

[0204]

[0205] Among them, Z gi is the normalized expression value of the g-th gene in the ith cell, Lgi is the expression value after improved Log transformation calculated previously, μ g is the weighted average Log2 expression value of gene g in all cells, σ g is the standard deviation based on the weights.

[0206] Set the previously calculated L gi is 3.635, μ g The calculation is as follows:

[0207]

[0208] Set L of N cells to be analyzed gi They are 3.635, 2.8 and 4.1 respectively, and the weight W g are 0.55, 0.6 and 0.65 respectively, calculate μ g as follows:

[0209]

[0210] μ g ≈3.16

[0211] Similarly, the standard deviation σ based on the weight g The calculation is as follows:

[0212]

[0213] After normalization, the normalized expression value Z of gene g in the i-th cell is gi It can be further used for model training and testing to ensure that the expression levels of different genes are more comparable and consistent.

[0214] Embodiment 4:

[0215] See also Figure 3 In this example, a set of single-cell transcriptome data is set, and the data set contains gene expression information of AD patients and healthy control samples. The goal is to identify genes with significant differences in these data to help understand the pathological mechanism of AD.

[0216] First, pseudo-count adjustment is required for gene expression data to eliminate noise and technical bias. In this process, a pseudo-count method based on Gamma distribution adjustment is used. The pseudo-count adjustment formula is as follows:

[0217]

[0218] Among them, C adjusted is the adjusted pseudo count, C original is the raw count, It is a Gamma function dynamically calculated based on the global characteristics of gene expression data and used to generate the adjustment coefficient. The parameters λ and k are the shape and scale parameters of the Gamma distribution, respectively.

[0219] Suppose there is a gene g, and its raw UMI count C in a cell original =100. To adjust this pseudo count, we need to determine the shape parameter λ and scale parameter k of the Gamma distribution. We set λ = 2.5 and k = 0.1 (these values ​​can be determined dynamically by analyzing the global characteristics of the data).

[0220] Next, calculate the adjustment coefficient for the Gamma distribution:

[0221]

[0222] Assuming the random number generated from the Gamma distribution is 25 (which can be obtained by calculation or simulation), the adjusted pseudo count is:

[0223] C adjusted =100+25=125

[0224] In this way, the adjusted pseudo count value C is obtained. adjusted =125.

[0225] Through this adjustment, the expression value of each gene in the data more accurately reflects its actual expression level in the cell and reduces the impact of technical noise.

[0226] Suppose that during the analysis, it is found that there are significant differences in pseudo count values ​​of certain genes between AD patients and healthy controls. For example, the adjusted pseudo count mean of gene g in AD patients is 150, while the mean in healthy controls is 90. Statistical tests can be used to confirm whether these differences are significant and identify these genes as AD markers.

[0227] Then, a composite normalization method based on a multivariate regression model was used to eliminate batch effects and cell size differences in single-cell data, as well as other deviations from experimental design and sample processing. It is assumed that pseudo counts have been adjusted through the Gamma distribution, and differential genes between AD patients and healthy controls have been identified. Next, these data need to be further composite normalized to improve the accuracy of the analysis.

[0228] Suppose there is a gene g, and its adjusted pseudo count in a certain cell is 125. At this time, it needs to be subjected to composite normalization. The formula for composite normalization is as follows:

[0229]

[0230] Among them, C normalizedis the normalized pseudo count, C adjusted is the adjusted pseudo count, β0, β1, and β2 are coefficients estimated by regression analysis, representing the effects of baseline correction, cell-specific factors, and batch effects, respectively. cell represents cell-specific factors, such as the effect of cell size; while g batch Represents batch effects, such as systematic deviations from different experimental batches.

[0231] It is assumed that through regression analysis, the following coefficient estimates are obtained: β0 = 0.5, β1 = 0.02, β2 = 0.1. In addition, it is assumed that in the current data set, the specific factor f of a cell cell is 2.0, and the batch effect factor g batch is 1.5.

[0232] First, calculate the value of the exponential term:

[0233] exp(β0+β1f cell +β2g batch )=exp(0.5+0.02×2.0+0.1×1.5)=exp(0.5+0.04+0.15)

[0234] =exp(0.69)≈1.993

[0235] Then, substitute the adjusted pseudo counts into the formula for normalization calculation:

[0236]

[0237] Through the normalization process, the original pseudo counts were adjusted from 125 to approximately 62.75, which can better eliminate the impact of batch effects and cell size differences, making the gene expression data between different samples more comparable.

[0238] Finally, EdgeR was used for differential expression analysis, and the dynamic local regression technique was introduced to optimize the estimation of dispersion parameters. In the single-cell transcriptome data, the data has been composite normalized, and EdgeR is now used for differential expression analysis to find genes that are significantly different between Alzheimer's disease (AD) patients and healthy controls.

[0239] In traditional differential expression analysis, the estimation of dispersion parameters affects the statistical significance assessment of gene expression. However, due to the high heterogeneity of single-cell transcriptome data, simple dispersion estimation methods can lead to inaccurate results. Therefore, the dynamic local regression (LOESS) technique is introduced to optimize the estimation of dispersion parameters.

[0240] Assume that there is a gene g, and its initial dispersion parameter estimate is 0.05. LOESS is needed to dynamically adjust it. The formula for dynamic local regression dispersion estimation is:

[0241] φa djusted =LOESS(φ g ,{x i},{y i},θ)

[0242] Among them, φ g is the initial dispersion parameter of gene g, {x i} represents the data point of gene expression, {y i} represents the observed dispersion data points and θ is the smoothing parameter.

[0243] Assume that in the data set, the expression of gene g is 1000, the observed dispersion is 0.05, and there are several data points of other genes: for example, the expression of gene a is 500, the dispersion is 0.08; the expression of gene b is 1500, the dispersion is 0.03. Based on these data points, LOESS will perform local weighted regression and dynamically select the smoothing parameter θ to optimize the estimation of the dispersion parameter.

[0244] Through preliminary analysis, the smoothing parameter θ is determined to be 0.75 (this is a common value that can be adjusted according to the local density and variability of the data). Under this setting, LOESS will adjust the dispersion of gene g based on nearby data points. After adjustment by LOESS, the dispersion parameter of gene g becomes:

[0245] φ adjusted =0.04

[0246] This adjusted dispersion parameter can better reflect the true variation of gene g, thereby improving the accuracy of differential expression analysis.

[0247] LOESS introduces data-driven smoothing parameter selection, making the estimation of dispersion parameters more flexible and accurate, especially for highly heterogeneous datasets such as single-cell transcriptome data. The adjusted dispersion parameter can not only more accurately reflect the true volatility of gene expression, but also enhance the stability and reliability of subsequent statistical analysis. For practical applications, the smoothing parameter θ is usually selected in the range of 0.5 to 1.0, and can be adjusted according to the specific characteristics of the data to ensure that the smoothing effect of LOESS is neither excessive nor insufficient.

[0248] Embodiment 5:

[0249] Combination Figure 4In this example, a large amount of single-cell transcriptome data from Alzheimer's disease (AD) patients and healthy controls has been collected and preprocessed. Now, these data are further analyzed through specific model training and optimization techniques to build a classification model that can effectively distinguish AD patients from healthy controls.

[0250] First, feature selection was performed using an improved differential expression analysis method. In this process, a significance evaluation method based on negative binomial distribution was used. Suppose a specific gene g is being analyzed, and the raw count in the i-th cell is C gi = 20. To assess the significance score of the gene in this cell, the following formula was used:

[0251]

[0252] In this formula, k and θ are the parameters of the negative binomial distribution. Assume that through previous statistical analysis, the range of values ​​of these parameters is estimated to be k = 5 and θ = 2. These values ​​can be substituted to calculate the significance score S gi .

[0253] Through calculation, we get:

[0254]

[0255] After further calculation, the significance score S is set to gi ≈0.75. This score indicates that in this cell, the expression of gene g is significant and can be selected as one of the features.

[0256] After feature selection, nonlinear dimensionality reduction techniques such as autoencoders are used to capture complex patterns in gene expression data. Autoencoders are neural networks that compress and reconstruct data. Its loss function is defined as:

[0257]

[0258] Among them, L AE is the loss function of the autoencoder, x i is the input data (e.g. part of the gene expression data after feature selection), f θ is the network function, θ(t) is the time-dependent regularization parameter, and λ is the regularization coefficient. We set λ = 0.01 as the regularization coefficient, which helps prevent overfitting while ensuring that the autoencoder can capture the complex patterns of the data.

[0259] In this example, it is assumed that the autoencoder is trained and the final loss function value converges to L AE = 0.05, which indicates that the model is able to reconstruct the input data well while taking the regularization constraints into account.

[0260] Then, a deep learning model is used to further improve the accuracy of Alzheimer's disease (AD) diagnosis. In order to effectively deal with the sparsity and cell-to-cell heterogeneity in single-cell transcriptome data, the structure of the deep learning model is adjusted, and the dropout technology is introduced to avoid model overfitting.

[0261] First, a deep neural network model is designed and a customized dropout mechanism is introduced to handle sparse data. The dropout ratio is adjusted according to the following function:

[0262]

[0263] Where D(x) is the output after dropout, and α and β are control parameters. In order to better adapt to the actual data, the range of α is set to [0.1, 0.5] and the range of β is set to [0.01, 0.1]. For the input node, x = 0.8, α = 0.3 and β = 0.05 are selected, and the calculation results are:

[0264]

[0265] Therefore, the output value of the node is reduced to 0.2 after dropout processing, which effectively avoids model overfitting and enhances the robustness of the model.

[0266] Next, in order to deal with the problem of class imbalance, a customized classification loss function is designed:

[0267]

[0268] Among them, L class is the classification loss function, w c is the class weight, y c and are the true and predicted labels, respectively, and μ is the smoothing regularization coefficient. Different weights are set for AD patients and healthy controls, and the weight of AD patients is set to w AD =0.7, the weight of the healthy control is w HC = 0.3. In addition, μ is set to 0.05 to suppress extreme fluctuations in the data.

[0269] During the training process, the output of the model at a certain moment is set to and The true label is y AD =1 and y HC = 0. The loss function at this time can be calculated as:

[0270]

[0271] The final loss function value is approximately 0.1138. Adding the smoothing term ensures that the model will not cause excessive errors due to the influence of the minority class.

[0272] Finally, the representativeness and generalization ability of the model are further improved through the clustering-based K-fold validation strategy. This strategy ensures that the samples in each fold are well representative during the cross-validation process, thereby enhancing the model's predictive ability on new data.

[0273] First, the samples are clustered using the following formula to calculate the similarity scores between samples:

[0274]

[0275] Among them, K(x i ,x j ) is the similarity score between samples i and j, and γ and η are adjustment parameters. In order to obtain a reasonable similarity score, the range of γ is 0.1, 1.0 and the range of η is 0.5, 2.0. Set γ = 0.5 and η = 1.0, for sample x i and x j The distance is |x i -x j | 2 =2.0, the similarity score is calculated as follows:

[0276]

[0277] Therefore, the sample x i and x j The similarity score between them is 0.622. This score helps ensure that the samples are representative in each fold, so that the model can effectively learn the structural characteristics of the data in each fold.

[0278] Next, multiple evaluation indicators were used to optimize the model parameters, and a comprehensive evaluation function was used to evaluate the performance of the model:

[0279]

[0280] Among them, Q is the comprehensive evaluation function, ω m is the weight of the mth evaluation index, φ m (t) and ψ m (t) are the time-dependent functions of the mth evaluation index. Three evaluation indexes are used: accuracy, sensitivity, and specificity, and their weights are set to ω Accuracy =0.4,ω Sensitivity =0.3 and ω Specificity=0.3. The time-dependent functions of each evaluation index are set as follows:

[0281] φ Accuracy (t) = 0.9·e -0.1t

[0282] ψ Accuracy (t) = 0.95·e -0.05t

[0283] φ Sensitivity (t) = 0.85·e -0.1t

[0284] ψ Sensitivity (t) = 0.9·e -0.05t

[0285] φ specificity (t) = 0.8 e -0.1t

[0286] ψ specificity (t) = 0.85·e -0.05t

[0287] The calculation results are as follows:

[0288] Q=0.4·∫0 1 (0.9·e -0.1t 0.95 e -0.05t )dt+0.3·∫0 1 (0.85·e -0.1t 0.9 e -0.05t )dt+0.3·∫0 1 (0.8·e -0.1 t·0.85·e -0.05t )dt

[0289] Q=0.4·∫0 1 0.855·e -0.15t dt+0.3·∫0 1 0.765 e -0.15t dt+0.3·∫0 1 0.68 e -0.15t dt

[0290]

[0291] Q≈0.4·5.7+0.3·5.1+0.3·4.53≈2.28+1.53+1.36=5.17

[0292] The result of the comprehensive evaluation function Q is 5.17, which reflects the comprehensive performance of the model in multiple evaluation indicators and provides a clear direction for model optimization.

[0293] The above shows and describes the basic principles, main features and advantages of the present invention. It should be understood by those skilled in the art that the present invention is not limited to the above embodiments, and the above embodiments and descriptions are only for explaining the principles of the present invention. Without departing from the spirit and scope of the present invention, the present invention may have various changes and improvements, which fall within the scope of the present invention to be protected. The scope of protection of the present invention is defined by the attached claims and their equivalents.

Claims

1. A method for single-cell transcriptome data-assisted AD analysis and classification, characterized in that The following steps are involved: S1. Sample collection and preparation: S1.

1. Obtain brain tissue samples from AD patients and age-matched healthy controls; use a brain tissue biobank or obtain samples through a brain tissue donation program; S1.

2. Use cell separation techniques including flow cytometry and FACS to process brain tissue and obtain single cell suspension; separate different types of brain cells including neurons, astrocytes, and microglia through mechanical and enzymatic methods; S2. Single-cell transcriptome sequencing: S2.

1. Sequence the isolated single cells using 10xGenomicsChromium's high-throughput single-cell transcriptome sequencing platform to obtain the gene expression profile of each cell. S2.

2. Perform data quality control to remove low-quality cells and low-expression genes; use CellRanger's data analysis tool to process the raw sequencing data and generate a gene expression matrix; S3. Data preprocessing: S3.

1. Standardize the gene expression data, including using Log2 transformation and Z-score normalization, to eliminate technical bias between different samples; S3.

2. Use dimensionality reduction algorithms including principal component analysis (PCA), t-SNE, and UMAP to reduce the dimensionality of high-dimensional data; S4. Feature extraction: S4.

1. Differential expression analysis including DESeq2 and EdgeR was used to identify genes with significant differences between AD patients and healthy controls. S4.

2. Gene set enrichment analysis was performed using GSEA and pathway analysis including KEGG and GO to identify key gene pathways and biological processes related to AD; S5. Classification and diagnosis of AD: S5.

1. Input the single-cell transcriptome data of the new sample into the trained classification model to perform AD classification prediction; S5.

2. Output the classification results and combine them with the patient's clinical data and other biomarkers including Aβ and tau protein levels to provide comprehensive auxiliary diagnosis recommendations.

2. The method for single-cell transcriptome data-assisted AD analysis and classification according to claim 1, characterized in that The data quality control method comprises: S1. Removal of low-quality cells: Calculate the total UMI, UniqueMolecularIdentifier count U, and number of detected genes G for each cell using the formula: Calculate cell quality score Q c , cells below the preset threshold are considered low-quality cells; Remove cells with too high mitochondrial gene expression ratio MGR; use the formula: Among them, M mito is the total UMI number of mitochondrial gene expression, M total is the total UMI number of all genes expressed in the cell; set a threshold θ, and remove the cell when MGR>θ; Low-quality cells that do not meet the standards are removed using cell filtering criteria including UMI counts, gene numbers, and mitochondrial gene ratios; the comprehensive score S is calculated as follows: Among them, α and β are weight parameters; cells below the set threshold are removed; S2. Removal of low-expression genes: Calculate the expression frequency F of each gene in all cells; use the formula: Among them, F g is the expression frequency of gene g, E gi is the expression level of gene g in the ith cell, N is the total number of cells; set the gene expression threshold δ to remove genes expressed in less than the percentage of δ cells; the removal condition is: if F g <δ, then remove gene g.

3. The method for single-cell transcriptome data-assisted AD analysis and classification according to claim 1, characterized in that The CellRanger software is used to process the original sequencing data FASTQ file to generate a gene expression matrix; The specific steps are as follows: S1. Read the original sequencing data and perform gene sequence alignment; CellRanger uses splicing alignment to align reads to the reference genome using the formula: Among them, A i is the alignment score of the i-th read, S i is the number of successfully aligned gene fragments, R i is the total number of reads; S2. Then, the UMI tags are deduplicated to generate redundant gene expression counts. The UMI deduplication formula is: Among them, U unique is the UMI count after redundancy removal, M is the number of all UMI tags, and N j is the repeat count of the jth UMI label; S3, generate the initial gene expression matrix, and record the expression level of each gene in each cell; The expression matrix generation formula is: Among them, E gi is the expression level of the g-th gene in the i-th cell, C gi is the UMI count of the g-th gene in the i-th cell, U total is the total UMI count in the i-th cell.

4. The method for single-cell transcriptome data-assisted AD analysis and classification according to claim 1, characterized in that The method for eliminating technical bias between different samples includes: S1. First, based on the Poisson distribution characteristics of the UMI counts of single-cell data, an improved Log transformation is performed on the UMI counts in each cell using the equation: Among them, L gi is the improved Log-transformed expression value of the g-th gene in the i-th cell, C gi is the original UMI count; by smoothing small UMI count values, the impact of noise is reduced; S2. Then before Z-score normalization, a weight factor is introduced based on the frequency of each gene in all cells, using the equation: Among them, W g is the weight of gene g, F g is the frequency of gene g in all cells, k and τ are parameters that adjust the shape of the curve; S2.

1. Data standardization using weighted Z-score standardization: Among them, Z gi is the normalized expression value of the g-th gene in the i-th cell; μ g is the weighted average Log2 expression value of gene g in all cells; σ g is the standard deviation based on the weights; N is the total number of cells; Among them, μ g and σ g are the weighted mean and standard deviation, respectively, and are calculated as follows: The contribution of each gene was adjusted by weighting the normalized gene expression frequency.

5. The method for single-cell transcriptome data-assisted AD analysis and classification according to claim 1, characterized in that The method for identifying genes with significant differences between AD patients and healthy controls was constructed as follows: S1. Use a pseudo-counting method based on Gamma distribution adjustment; adjust each count according to the potential distribution characteristics of gene expression, and improve the pseudo-counting formula: Where: γ is the Gamma function used to generate the adjustment coefficient; λ and k are shape and scale parameters dynamically calculated based on the global characteristics of gene expression data; S2. Due to batch effects and differences in cell size in single-cell data, a composite normalization method based on a multivariate regression model was used to adjust for multiple deviations from experimental design and sample processing; the composite normalization formula was used: Among them: β0, β1, β2 are coefficients estimated by regression analysis, representing the effects of baseline correction, cell-specific factors, and batch effects, respectively; S3. When using EdgeR for differential expression analysis, the estimation of dispersion parameters is optimized by introducing dynamic local regression technology. The dynamic local regression dispersion estimation formula is: f adjusted =LOESS(φ g ,{x i },{y i },i) Where: LOESS represents the local weighted scatter point smoothing function; {x i },{y i } represent the gene expression and observed dispersion data points, respectively; θ is a dynamically selected smoothing parameter that is adjusted based on the local density and variability of the data.

6. Construction of a model for single-cell transcriptome data-assisted AD analysis and classification, characterized by The following steps are involved: S1. Select machine learning algorithms including random forest, support vector machine, and neural network to build a classification model; S2, using the extracted differential gene expression data as feature input for model training; using K-fold cross validation to optimize model parameters; S3. Verify the model performance through an independent test set and evaluate the accuracy, sensitivity, specificity and stability of the model; use indicators including AUC-ROC curve and confusion matrix to evaluate the model.

7. The method for single-cell transcriptome data-assisted AD analysis and classification according to claim 6, characterized in that The model training and model parameter optimization method includes: S1. First, an improved differential expression analysis method was used, using feature selection based on negative binomial distribution, where the significance evaluation formula is: Among them, S gi is the significance score of the g-th gene in the i-th cell, C gi is the raw count, k and θ are the parameters of the negative binomial distribution; S1.

1. Simultaneously combine nonlinear dimensionality reduction techniques including autoencoders to capture complex patterns in gene expression data. The loss function of the autoencoder is: Among them, L AE is the loss function of the autoencoder, x i is the input data, f θ is the network function, θ(t) is the time-dependent regularization parameter, and λ is the regularization coefficient; S2. A deep learning model is used and its structure is adjusted to adapt to the single-cell data characteristics including sparsity and intercellular heterogeneity. Model optimization is achieved through dropout technology, and the dropout ratio is controlled by the following function: Where D(x) is the output after dropout, α and β are control parameters; S2.

1. Use a customized loss function to solve the problem of class imbalance. The loss function is: Among them, L class is the classification loss function, w c is the class weight, y c and are the true and predicted labels respectively, μ is the smoothing regularization coefficient; S3. Finally, the K-fold validation strategy based on clustering is implemented to ensure the representativeness and generalization ability of each compromise. The clustering division is performed by the following formula: Among them, K(x i ,x j ) is the similarity score between samples i and j, γ and η are adjustment parameters; S3.

1. Optimize model parameters by combining multiple evaluation indicators at the same time. The comprehensive evaluation function is: Among them, Q is the comprehensive evaluation function, ω m is the weight of the mth evaluation index, φ m (t) and ψ m (t) are the time correlation functions of the mth evaluation index.

Images