Cell type annotation methods and annotation systems
Patent Information
- Application Number
- CN202510922198.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-07-04
- Publication Date
- 2026-09-01
- Estimated Expiration
- 2045-07-04
AI Technical Summary
[0003]但是，现有的空间转录组数据的得到每一个样本中的细胞类型组成中，对注释细胞类型的准确度较低
[0044]The beneficial effects of this invention are as follows: First, common genes between single-cell and spatial transcriptome data are identified and their corresponding expressions are extracted. A common encoder maps these gene expressions to two latent features: a first latent feature and a second latent feature. By aligning the feature distributions of the first and second latent features in a common space, data domain differences and batch effects are eliminated, making the feature distributions of single-cell data and spatial transcriptome data comparable in the same semantic space. Finally, based on a preset distance function, known cell type labels from single-cell data are accurately transferred to spatial transcriptome data through mapping relationships. These steps work together to ensure the effectiveness of features and consistency across datasets during annotation, thereby improving annotation accuracy.
Smart Images
Figure CN120932748B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of biological detection technology, specifically to a cell type annotation method and annotation system. Background Technology
[0002] Spatial transcriptome data samples obtained based on imaging technology have both gene expression counts and spatial coordinates at single-cell resolution, which can better establish a single-cell level correspondence with single-cell RNA (ribonucleic acid) sequencing data from the same tissue, and transfer the cell type of single-cell data to spatial transcriptome data.
[0003] However, existing spatial transcriptome data, which derives cell type composition from each sample, have low accuracy in annotating cell types. Summary of the Invention
[0004] To address the shortcomings of existing technologies, this invention provides a cell type annotation method that can effectively improve the accuracy of cell type annotation.
[0005] This application provides a cell type annotation method, the annotation method comprising:
[0006] Identify common genes in single-cell data and spatial transcriptome data, and extract the expression of the first gene corresponding to the common gene in the single-cell data and the expression of the second gene corresponding to the common gene in the spatial greening group data;
[0007] A first latent feature is obtained by extracting features from the expression of the first gene using a common encoder, and a second latent feature is obtained by extracting features from the expression of the second gene.
[0008] Align the distributions of the first latent feature and the second latent feature in a common space;
[0009] Cell type annotation is performed on the spatial transcriptome data based on a preset distance function, combined with the first latent feature and the second latent feature.
[0010] In one aspect, the steps of identifying common genes in single-cell data and spatial transcriptome data, and extracting the expression of a first gene corresponding to the common gene in the single-cell data and the expression of a second gene corresponding to the common gene in the spatial greening data, include:
[0011] Obtain first raw omics data from single-cell data and second raw omics data from spatial transcriptome data from the same tissue, and identify common genes in the single-cell data and the spatial transcriptome data;
[0012] The first raw omics data and the second raw omics data are filtered and preprocessed respectively to obtain the first gene expression corresponding to the common gene of the single cell data and the second gene expression corresponding to the common gene of the spatial greening group data.
[0013] In one aspect, the steps of filtering and preprocessing the first raw omics data and the second raw omics data respectively to obtain the expression of a first gene corresponding to the single-cell data and its common gene, and the expression of a second gene corresponding to the spatially converted green group data and its common gene, include:
[0014] Low-quality genes and cells were filtered from the first and second raw omics data, respectively.
[0015] Genes with expression levels below a first threshold and cells with expression levels below a second threshold are removed from the original omics data, where the second threshold is greater than the first threshold.
[0016] Highly expressed differentially expressed genes are extracted from the first and second original omics data. The highly expressed differentially expressed genes are then normalized and logarithmically transformed to obtain the expression of the first and second genes.
[0017] In one aspect, the steps of extracting features from the expression of the first gene using a co-encoder to obtain a first latent feature and extracting features from the expression of the second gene to obtain a second latent feature include:
[0018] The first gene expression and the second gene expression are input into a common encoder to obtain the first latent feature and the second latent feature;
[0019] Among them, X SC For first gene expression, X ST For the expression of the second gene, Z SC Z is the first potential feature. ST This is the second potential feature;
[0020] Z SC =MLP1(Attention(MLP2(X) SC )));
[0021] Z ST =MLP1(Attention(MLP2(X) ST MLP1 and MLP2 are multilayer perceptrons, and Attention is an attention machine.
[0022] In one aspect, prior to the step of aligning the distributions of the first latent feature and the second latent feature in a common space, the method includes:
[0023] The first latent feature is input into the classifier, and supervised learning is performed using known cell type labels to optimize the classification accuracy of the cell type of the first latent feature.
[0024] In this context, the corresponding loss function for supervised learning is the cross-entropy loss function.
[0025] K represents the number of cell types in the single-cell data, y k p represents the one-hot encoded vector representing the true label. k This represents the probability value predicted by the classifier for each category.
[0026] In one aspect, after the step of aligning the distributions of the first latent feature and the second latent feature in a common space, the method includes:
[0027] The first latent feature is reconstructed based on the first decoder to form the first reconstructed data, and the second latent feature is reconstructed based on the second decoder to form the second reconstructed data;
[0028] The first decoder includes two multilayer perceptrons and one attention mechanism layer, and the second decoder includes three multilayer perceptrons.
[0029] For the first reconstructed data, For the second reconstructed data, φ sc For the first decoder, φ ST This is the second decoder.
[0030] In one aspect, the reconstruction loss functions of the first decoder and the second decoder are:
[0031]
[0032] n represents the number of cells input for single-cell ribonucleic acid data, and m represents the number of cells input for spatial transcriptome data. For single-cell RNA data, the input gene expression of the i-th cell is... This represents gene expression after reconstruction of the i-th cell from single-cell ribonucleic acid data. For the input gene expression of the j-th cell in the spatial transcriptome data, This represents the gene expression in the j-th cell after reconstruction of the spatial transcriptome data.
[0033] In one aspect, the step of annotating the spatial transcriptome data by cell type based on a preset distance function combined with the first latent feature and the second latent feature includes:
[0034] The similarity between the first latent feature and the second latent feature is measured based on a preset distance function, and a mapping relationship is constructed.
[0035] Cell type annotation is performed on the spatial transcriptome data using the mapping relationship.
[0036] In one aspect, the step of aligning the distributions of the first latent feature and the second latent feature in a common space includes:
[0037] The first and second latent features are fed into a domain classifier for adversarial learning; wherein, the domain alignment loss function in adversarial learning is the binary cross-entropy.
[0038] in This represents the true domain label of sample i. The domain classifier predicts the probability that sample i belongs to the source domain.
[0039] To address the aforementioned problems, this application also provides a cell type annotation system, the annotation system comprising:
[0040] The extraction module is used to identify common genes in single-cell data and spatial transcriptome data, and to extract the expression of the first gene corresponding to the common gene in the single-cell data and the expression of the second gene corresponding to the common gene in the spatial greening group data.
[0041] The extraction module is further configured to use a common encoder to extract features from the expression of the first gene to obtain a first latent feature, and to extract features from the expression of the second gene to obtain a second latent feature;
[0042] An alignment module is used to align the distributions of the first latent feature and the second latent feature in a common space;
[0043] The annotation module annotates the spatial transcriptome data by cell type based on a preset distance function combined with the first latent feature and the second latent feature.
[0044] The beneficial effects of this invention are as follows: First, common genes between single-cell and spatial transcriptome data are identified and their corresponding expressions are extracted. A common encoder maps these gene expressions to two latent features: a first latent feature and a second latent feature. By aligning the feature distributions of the first and second latent features in a common space, data domain differences and batch effects are eliminated, making the feature distributions of single-cell data and spatial transcriptome data comparable in the same semantic space. Finally, based on a preset distance function, known cell type labels from single-cell data are accurately transferred to spatial transcriptome data through mapping relationships. These steps work together to ensure the effectiveness of features and consistency across datasets during annotation, thereby improving annotation accuracy. Attached Figure Description
[0045] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the accompanying drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. In all the drawings, similar elements or parts are generally identified by similar reference numerals. In the drawings, the elements or parts are not necessarily drawn to scale.
[0046] Figure 1 This is a schematic diagram illustrating the steps of the cell type annotation method in this application;
[0047] Figure 2 For this application Figure 1 A schematic diagram of the process steps for extracting the first and second expressed genes;
[0048] Figure 3 For this application Figure 2 A schematic diagram illustrating the process steps for filtering and preprocessing the first and second raw omics data.
[0049] Figure 4 This is a schematic diagram of the process steps for obtaining the first and second latent features in this application;
[0050] Figure 5 This is a schematic diagram of the workflow steps for supervised learning using known cell type labels in this application;
[0051] Figure 6 This is a schematic diagram of the process steps for forming the first reconstructed data and the second reconstructed data in this application;
[0052] Figure 7 This is a schematic diagram illustrating the steps for annotating cell types in this application;
[0053] Figure 8 This is a schematic diagram illustrating the steps in this application of putting the first latent feature and the second latent feature into a domain classifier for adversarial learning.
[0054] Figure 9 This is a schematic diagram of the cell type annotation system of this application;
[0055] Figure 10 This is a flowchart of the multimodal learning model for integrating and analyzing single-cell data and spatial transcriptome data in this application;
[0056] Figure 11 This diagram illustrates a comparison of the cell type annotation method proposed in this application with other existing methods.
[0057] Figure descriptions: 100, Extraction module; 200, Alignment module; 300, Reconstruction module; 400, Annotation module. Detailed Implementation
[0058] The embodiments of the technical solution of the present invention will now be described in detail with reference to the accompanying drawings. These embodiments are merely illustrative of the technical solution of the present invention and are therefore intended to limit the scope of protection of the present invention.
[0059] It should be noted that, unless otherwise stated, the technical or scientific terms used in this application should have the ordinary meaning as understood by one of ordinary skill in the art to which this invention pertains.
[0060] like Figure 1 As shown, this application provides a cell type annotation method, which includes:
[0061] Step S10: Identify common genes in single-cell data and spatial transcriptome data; extract the expression of the first gene corresponding to the common gene in single-cell data and the expression of the second gene corresponding to the common gene in spatial greening data; by screening common genes in single-cell data and spatial transcriptome data and preprocessing the raw data of single-cell data and spatial transcriptome data, ensure that the input first gene expression and second gene expression data have biological consistency, remove technical noise and invalid information, provide a reliable data foundation for subsequent feature extraction, and avoid annotation bias caused by differences in data quality.
[0062] Step S20: Use a common encoder to extract features from the expression of the first gene to obtain the first latent feature, and use a common encoder to extract features from the expression of the second gene to obtain the second latent feature; use a common encoder containing a multilayer perceptron (MLP) and an attention mechanism to extract features from the preprocessed gene expression data, and use the attention mechanism to focus on key gene features related to cell type, suppress irrelevant noise, and map high-dimensional gene expression to low-dimensional latent features, so that the extracted first and second latent features are more cell type discriminative and enhance the ability of features to represent cell type.
[0063] Step S30: Align the distributions of the first and second latent features in the common space; through adversarial learning and supervised learning using a domain classifier, align the latent feature distributions of single-cell data and spatial transcriptome data in the common space, eliminate batch effects and domain differences between different datasets, make the features of the two types of data comparable in the same semantic space, ensure that the known cell type feature distributions in single-cell data can be accurately transferred to spatial transcriptome data, and improve the accuracy of cross-dataset feature mapping.
[0064] Step S40: Cell type annotation is performed on the spatial transcriptome data based on a preset distance function combined with the first and second latent features. In the aligned latent feature space, the preset distance function is used to measure the feature similarity between the spatial transcriptome data and the single-cell data. A mapping relationship is constructed by calculating the feature distance, and the precise cell type labels in the single-cell data are assigned to the spatial transcriptome cells according to their similarity. Since the feature distribution is aligned, the distance measurement can effectively reflect the true association of cell types, thereby achieving high-precision cell type annotation.
[0065] In this embodiment, common genes between single-cell and spatial transcriptome data are first identified and their corresponding expressions are extracted. A common encoder maps these gene expressions to two latent features: a first latent feature and a second latent feature. By aligning the feature distributions of the first and second latent features in the common space, data domain differences and batch effects are eliminated, making the feature distributions of single-cell data and spatial transcriptome data comparable in the same semantic space. Finally, based on a preset distance function, known cell type labels from the single-cell data are accurately transferred to the spatial transcriptome data through mapping relationships. These steps work together to ensure the effectiveness of features during annotation and consistency across datasets, thereby improving annotation accuracy.
[0066] like Figure 2 As shown, the steps for identifying common genes in single-cell data and spatial transcriptome data, extracting the expression of the first gene corresponding to the common gene in single-cell data, and the expression of the second gene corresponding to the common gene in spatial greening data, include:
[0067] Step S110 involves acquiring the first raw omics data from single-cell data and the second raw omics data from spatial transcriptome data from the same tissue, and identifying common genes between the single-cell and spatial transcriptome data. It also involves acquiring the first raw omics data from single-cell data and the second raw omics data from spatial transcriptome data from the same tissue, and using gene intersection analysis to identify the common gene set. This step ensures that both types of data are based on the same biological marker system, laying a consistent foundation for subsequent cross-dataset feature mapping and avoiding feature shifts caused by differences in gene sets.
[0068] Step S120 involves filtering and preprocessing the first and second raw omics data to obtain the first gene expression corresponding to the common genes in single-cell data, and the second gene expression corresponding to the common genes in spatial greening data. Low-quality genes and cells are filtered for both types of raw omics data, resulting in the first gene expression matrix for single-cell data and the second gene expression matrix for spatial transcriptome data. This preprocessing process ensures the reliability and cross-dataset comparability of gene expression data by eliminating technical noise and standardizing expression levels.
[0069] like Figure 3 As shown, the steps for filtering and preprocessing the first and second raw omics data to obtain the expression of the first gene corresponding to the common gene in single-cell data and the expression of the second gene corresponding to the common gene in spatially converted greening group data include:
[0070] Step S121 involves filtering low-quality genes and cells from the first and second raw omics data, respectively; and performing low-quality gene and cell filtering operations on the first raw omics data of single-cell data and the second raw omics data of spatial transcriptomics data, respectively. By setting basic screening conditions, invalid genes and cells in the data due to poor sequencing quality, low biological activity, etc., are removed, noise interference is reduced, and subsequent analyses are ensured to be based on reliable biological data.
[0071] Step S122: Remove genes with expression levels below the first threshold and cells with gene expression levels below the second threshold from the original omics data. The second threshold is greater than the first threshold. Further purify the data using a dual threshold standard: at the gene level, remove genes with expression levels below the "first threshold" to avoid interference from low-expression genes in subsequent analyses; at the cell level, remove cells with gene expression levels below the "second threshold" (the second threshold is greater than the first threshold) to ensure that the remaining cells have sufficient gene expression activity and avoid data bias caused by cell damage or insufficient sequencing depth.
[0072] Step S123: Extract highly expressed differentially expressed genes from the first and second raw omics data. Normalize and logarithmically transform these genes to obtain the first and second gene expression levels. Extract genes that show high expression and significant differences in both types of omics data from the filtered raw data. Normalize these genes to eliminate expression biases caused by sequencing depth, sample preparation, and other technical factors. Then, perform a logarithmic transformation to stabilize data variance and compress numerical ranges. Finally, obtain the first gene expression matrix from single-cell data and the second gene expression matrix from spatial transcriptome data, providing standardized input data for subsequent feature extraction.
[0073] like Figure 4As shown, the steps of extracting features from the expression of a first gene to obtain a first latent feature and extracting features from the expression of a second gene to obtain a second latent feature using a common encoder include:
[0074] Step S210: Input the first gene expression and the second gene expression into a common encoder to obtain the first latent feature and the second latent feature. The gene expression can be in matrix form. Input the preprocessed single-cell data first gene expression matrix and the spatial transcriptome data second gene expression matrix into a common encoder composed of a multilayer perceptron (MLP) and an attention mechanism. Through hierarchical computation of the encoder, the first layer MLP first performs preliminary feature transformation on the gene expression data, then the attention mechanism focuses on key gene features related to cell type and suppresses irrelevant noise, and finally the second layer MLP maps the high-dimensional gene expression into low-dimensional first latent features and second latent features, realizing feature representation of the two types of data in a common semantic space, providing a discriminative and robust feature representation for subsequent distribution alignment and cell type annotation.
[0075] Among them, X SC For first gene expression, X ST For the expression of the second gene, Z SC Z is the first potential feature. ST Z is the second potential feature; SC =MLP1(Attention(MLP2(X) SC ))).
[0076] Z ST =MLP1(Attention(MLP2(X) ST MLP1 and MLP2 are multilayer perceptrons, and Attention is an attention machine.
[0077] like Figure 5 As shown, prior to the step of aligning the distributions of the first latent feature and the second latent feature in the common space, the procedure includes:
[0078] Step S50: Input the first latent feature into the classifier and perform supervised learning using known cell type labels to optimize the classification accuracy of the cell type of the first latent feature; input the first latent feature extracted from single-cell data by the common encoder into the classifier and perform supervised learning training using known cell type labels. Calculate the deviation between the predicted label and the true label using the cross-entropy loss function, and optimize the parameters of the classifier and the common encoder through backpropagation. This makes the feature distribution of different cell types in the first latent feature more discriminative, that is, the features of the same type of cell are highly aggregated, and the features of different types of cells are significantly separated, thereby improving the classification accuracy of cell types in single-cell data and providing a more accurate single-cell feature benchmark for feature alignment and label transfer of subsequent spatial transcriptome data.
[0079] In this context, the corresponding loss function for supervised learning is the cross-entropy loss function.
[0080] K represents the number of cell types in the single-cell data, y k p represents the one-hot encoded vector representing the true label. k This represents the probability value predicted by the classifier for each category.
[0081] like Figure 6 As shown, after the step of aligning the distributions of the first and second latent features in the common space, the process includes:
[0082] Step S60: Based on the first decoder, the first latent feature is reconstructed to form the first reconstructed data, and based on the second decoder, the second latent feature is reconstructed to form the second reconstructed data. The first decoder includes two layers of multilayer perceptron and one layer of attention mechanism, and the second decoder includes three layers of multilayer perceptron. The first decoder consists of two layers of multilayer perceptron and one layer of attention mechanism. After the first latent feature is input, the expression weights of key gene features are reweighted through the attention mechanism, and then mapped back to the high-dimensional space layer by layer through MLP to form the first reconstructed data, ensuring that the features of single-cell data can still restore the original gene expression pattern after dimensionality reduction.
[0083] The second decoder contains a three-layer multilayer perceptron, which processes the second latent features through a deeper network structure, restores the gene expression distribution of spatial transcriptome data layer by layer, and forms the second reconstructed data, which is adapted to the gene expression patterns and noise characteristics unique to spatial transcriptome data.
[0084] Step S60 optimizes the reconstruction loss function to ensure that key biological information is not lost during dimensionality reduction and alignment of latent features, thus providing a reliable guarantee for subsequent cell type annotation.
[0085] For the first reconstructed data, For the second reconstructed data, φ sc For the first decoder, φ ST This is the second decoder.
[0086] Furthermore, the reconstruction loss function of the first decoder and the second decoder is:
[0087]
[0088] n represents the number of cells input for single-cell ribonucleic acid data, and m represents the number of cells input for spatial transcriptome data. For single-cell RNA data, the input gene expression of the i-th cell is... This represents gene expression after reconstruction of the i-th cell from single-cell ribonucleic acid data. For the input gene expression of the j-th cell in the spatial transcriptome data, This represents the gene expression in the j-th cell after reconstruction of the spatial transcriptome data.
[0089] like Figure 7 As shown, the steps for annotating cell types in spatial transcriptome data based on a preset distance function combined with first and second latent features include:
[0090] Step S410: The similarity between the first and second latent features is measured based on a preset distance function, and a mapping relationship is constructed. Based on the latent features aligned in the common space, a preset distance function, such as Euclidean distance or cosine similarity, is used to calculate the similarity between the first latent feature of the single-cell data and the second latent feature of the spatial transcriptome data. By quantifying the spatial distance between the two types of features, a mapping relationship from spatial transcriptome features to single-cell features is established. That is, the feature vector of each spatial transcriptome cell is correlated with the similarity of the features of each cell type in the single-cell data, providing a quantitative basis for subsequent label transfer.
[0091] Step S420 involves annotating cell types in the spatial transcriptome data using a mapping relationship. Utilizing the feature mapping relationship constructed in S410, known cell type labels from single-cell data are transferred to the spatial transcriptome data. Specifically, for the second latent feature of each spatial transcriptome cell, the cell type label corresponding to the feature with the highest similarity in the single-cell data is found through the mapping relationship and directly assigned to that spatial cell, thereby completing the cell type annotation of the spatial transcriptome data and achieving accurate transfer of cell type information across datasets.
[0092] like Figure 8 As shown, the steps for aligning the distributions of the first and second latent features in a common space include:
[0093] Step S310: Input the first latent feature and the second latent feature into the domain classifier for adversarial learning; wherein, the domain alignment loss function in adversarial learning is the binary cross-entropy. The first latent features of single-cell data and the second latent features of spatial transcriptome data are input into a domain classifier, and the feature distributions of the two types of data are aligned through an adversarial learning mechanism. The goal of the domain classifier is to determine the data domain to which the features belong, while the co-encoder attempts to generate features that cannot be distinguished by the domain classifier. Through the game between the two, i.e., the "adversarial" interaction, the features are forced to learn cross-domain invariant biological characteristics, eliminating data domain differences such as batch effects.
[0094] In adversarial learning, binary cross-entropy is used as the domain alignment loss function. Its core principle is: if the domain classifier can accurately distinguish the source of features, the loss value is high; if the features generated by the encoder are indistinguishable by the domain classifier, the loss value is low. Through backpropagation optimization, the encoder is forced to learn "domain-invariant" features, meaning that regardless of whether the features come from single-cell data or spatial transcriptome data, their distribution in the common space tends to be consistent. This lays the foundation for feature consistency in subsequent cross-domain transfer of cell type labels, significantly improving annotation accuracy.
[0095] in This represents the true domain label of sample i. The domain classifier predicts the probability that sample i belongs to the source domain.
[0096] like Figure 9 As shown, this application also provides a cell type annotation system, which includes an extraction module 100, an alignment module 200, and an annotation module 400.
[0097] The extraction module 100 is used to identify common genes in single-cell data and spatial transcriptome data, and to extract the first gene expression corresponding to the common genes in single-cell data and the second gene expression corresponding to the common genes in spatial greening data. The extraction module is also used to extract features from the first gene expression using a common encoder to obtain the first latent feature and extract features from the second gene expression to obtain the second latent feature. The extraction module is mainly responsible for data preprocessing and feature extraction: First, the raw omics data of single-cell data and spatial transcriptome data from the same tissue are acquired. The common gene set is screened through gene intersection, and the raw data is preprocessed by low-quality gene and cell filtering, normalization, and logarithmic transformation to obtain the first gene expression matrix of single-cell data and the second gene expression matrix of spatial transcriptome data. Then, the common encoder composed of multilayer perceptron (MLP) and attention mechanism is used to extract features from the above gene expression matrix to generate the first latent feature of single-cell data and the second latent feature of spatial transcriptome data, providing standardized feature representations for subsequent analysis.
[0098] Alignment module 200 is used to align the distribution of the first and second latent features in the common space. The alignment module achieves feature distribution alignment through domain-adaptive adversarial learning: the first and second latent features are input into the domain classifier, and adversarial training is performed using the binary cross-entropy loss function. The domain classifier attempts to distinguish the data domain to which the features belong (single-cell domain or spatial transcriptome domain), while the co-encoder optimizes the parameters through the gradient inversion layer, forcing the generated features to be indistinguishable by the domain classifier, thereby eliminating the batch effect and domain difference between the two types of data, making the latent features consistent in distribution in the common space, and laying the foundation for cell type mapping across datasets.
[0099] The annotation module 400 annotates cell types in spatial transcriptome data based on a preset distance function combined with first and second latent features. The annotation module achieves cell type label transfer based on feature similarity: in the aligned latent space, a preset distance function (such as Euclidean distance) is used to measure the similarity between the second latent features of the spatial transcriptome data and the first latent features of the single-cell data, constructing a feature mapping relationship; by calculating the distance between the spatial transcriptome cell features and the centers of each cell type feature in the single cell, the label of the single cell corresponding to the smallest distance is assigned to that spatial cell, completing the cell type annotation of the spatial transcriptome data and achieving accurate label transfer across datasets.
[0100] The annotation system also includes a reconstruction module 300. The reconstruction module 300 reconstructs the first latent features into first reconstructed data based on a first decoder, and reconstructs the second latent features into second reconstructed data based on a second decoder. The reconstruction module preserves feature biological information through a decoder network: it uses the first and second decoders to reconstruct gene expression of the first and second latent features respectively, generating first and second reconstructed data; and it optimizes the reconstruction loss function through mean squared error to ensure that key biological features are not lost during dimensionality reduction and alignment, verifying the effectiveness of feature representation and providing reliable assurance for cell type annotation.
[0101] like Figure 10 The diagram shows a multimodal learning model for integrating single-cell RNA sequencing (SC) and spatial transcriptomics (ST) data. The core of this model is to achieve cell type annotation across modal data through feature encoding, adversarial alignment, classification, and reconstruction. The breakdown is as follows:
[0102] Step 1: Data Encoding and Adversarial Alignment
[0103] Input data:
[0104] X SC Includes gene expression matrix and known cell type labels.
[0105] XST Gene expression matrix of the same organization.
[0106] Shared Encoder: Using a multilayer perceptron, for X... SC X ST The common gene expression matrix encodes and generates low-dimensional latent features:
[0107] Z SC Preserve the cell type distinguishability of SC data.
[0108] Z ST Gene expression in compressed ST data, and Z SC Align to the same latent space.
[0109] Domain Classifier: Uses MLP to determine whether latent features come from the SC or ST domain, aiming to maximize the domain classification loss (binary cross-entropy); while Shared Encoder minimizes the domain classification loss through adversarial training, allowing Z to... SC Z ST The potential spatial distribution is "indistinguishable", eliminating data modal differences.
[0110] Classifier (cell type classifier): Using MLP to classify Z SC By training with known labels, optimize the classification loss (cross-entropy) to let Z SC It can accurately distinguish cell types, providing "reference labels" for subsequent ST annotation.
[0111] Decoders:
[0112] SC Decoder: Reconstructing Z using MLP SC for Calculate the reconstruction loss (mean squared error) to ensure that the encoding preserves the expression information of the SC gene.
[0113] ST Decoder: Similarly, refactor Z ST for The biological significance of calculating the protection of ST characteristics.
[0114] Step 2: Spatial Transcriptome Cell Type Annotation
[0115] In the latent space, Z SC Cell type features have been learned through a classifier (e.g., different colored triangles represent known types), Z ST (Circular) then uses a distance metric (such as Euclidean distance) to find Z. SCThe nearest cell type label (Transfer Label) is used to achieve cell type "transfer annotation" in ST data, thus solving the problem of missing labels in ST data.
[0116] By aligning cross-modal features through shared encoding, eliminating domain differences through adversarial learning, learning cell types through classifiers, and ensuring feature effectiveness through reconstruction loss, the labels of single-cell data are ultimately used to "migrate" and annotate spatial transcriptome data, thereby achieving integrated analysis of multimodal omics data.
[0117] like Figure 11 The image shows a bar chart comparing the performance of different cell type annotation methods on the STARmap dataset. The effectiveness of eight methods, including CTMAP, Tangram, and SPANN, is evaluated using three metrics: Accuracy, NMI (Normalized Mutual Information), and ARI (Adjusted Rand Index). The key interpretations are as follows:
[0118] 1. Meaning of the indicator
[0119] Accuracy: The ratio of predicted cell types to true labels; a higher accuracy indicates more accurate annotation.
[0120] NMI (Normalized Mutual Information): Measures the degree of information overlap between predicted labels and true labels; the higher the value, the stronger the consistency.
[0121] ARI (Adjusted Rand Index): Quantifies the degree of agreement between predicted clusters (cell type grouping) and the actual clusters. The higher the ARI, the closer the grouping is to the actual clustering.
[0122] 2. Conclusions of Method Comparison
[0123] CTMAP (red): Significantly outperforms in Accuracy, NMI, and ARI, making it the cell type annotation method proposed in this application. In cell type annotation tasks using spatial transcriptome data, the CTMAP method presented in this application exhibits the best overall performance.
[0124] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and they should all be covered within the scope of the claims and specification of the present invention.
Claims
1. A method for annotating cell types, characterized in that, The annotation methods include: Identify common genes in single-cell data and spatial transcriptome data, and extract the expression of the first gene corresponding to the common gene in the single-cell data and the expression of the second gene corresponding to the common gene in the spatial greening group data; A first latent feature is obtained by extracting features from the expression of the first gene using a common encoder, and a second latent feature is obtained by extracting features from the expression of the second gene. Align the distributions of the first latent feature and the second latent feature in a common space; Cell type annotation is performed on the spatial transcriptome data based on a preset distance function combined with the first latent feature and the second latent feature; The steps of extracting features from the expression of the first gene using a common encoder to obtain a first latent feature, and extracting features from the expression of the second gene to obtain a second latent feature, include: The first gene expression and the second gene expression are input into a common encoder to obtain the first latent feature and the second latent feature; Wherein, XSC represents the first gene expression, XST represents the second gene expression, ZSC represents the first latent trait, and ZST represents the second latent trait; ZSC=MLP1 (Attention(MLP2 (XSC))); ZST = MLP1 (Attention(MLP2 (XST))), where MLP1 and MLP2 are multilayer perceptrons, and Attention is an attention mechanism; The common encoder consists of a multilayer perceptron and an attention mechanism. Through the hierarchical computation of the encoder, the first layer MLP performs preliminary feature transformation on the gene expression data. Then, the attention mechanism focuses on the key gene features related to cell type and suppresses irrelevant noise. Finally, the second layer MLP maps the high-dimensional gene expression into low-dimensional first and second latent features, realizing the feature representation of the two types of data in the common semantic space. Prior to the step of aligning the distributions of the first latent feature and the second latent feature in a common space, the method includes: The first latent feature is input into the classifier, and supervised learning is performed using known cell type labels to optimize the classification accuracy of the cell type of the first latent feature. In this context, the corresponding loss function for supervised learning is the cross-entropy loss function. ; K represents the number of cell types in the single-cell data. The one-hot encoded vector representing the actual label. This represents the probability value predicted by the classifier for each category.
2. The annotation method according to claim 1, characterized in that, The steps of identifying common genes in single-cell data and spatial transcriptome data, and extracting the expression of the first gene corresponding to the common gene in the single-cell data and the expression of the second gene corresponding to the common gene in the spatial greening group data, include: Obtain first raw omics data from single-cell data and second raw omics data from spatial transcriptome data from the same tissue, and identify common genes in the single-cell data and the spatial transcriptome data; The first raw omics data and the second raw omics data are filtered and preprocessed respectively to obtain the first gene expression corresponding to the common gene of the single cell data and the second gene expression corresponding to the common gene of the spatial greening group data.
3. The annotation method according to claim 2, characterized in that, The steps of filtering and preprocessing the first and second raw omics data respectively to obtain the expression of the first gene corresponding to the common gene of the single-cell data and the second gene corresponding to the common gene of the spatially converted green group data include: Low-quality genes and cells were filtered from the first and second raw omics data, respectively. Genes with expression levels below a first threshold and cells with expression levels below a second threshold are removed from the original omics data, where the second threshold is greater than the first threshold. Highly expressed differentially expressed genes are extracted from the first and second original omics data. The highly expressed differentially expressed genes are then normalized and logarithmically transformed to obtain the expression of the first and second genes.
4. The annotation method according to claim 1, characterized in that, Following the step of aligning the distributions of the first latent feature and the second latent feature in a common space, the method includes: The first latent feature is reconstructed based on the first decoder to form the first reconstructed data, and the second latent feature is reconstructed based on the second decoder to form the second reconstructed data; The first decoder includes two multilayer perceptrons and one attention mechanism layer, and the second decoder includes three multilayer perceptrons. , For the first reconstructed data, For the second reconstructed data, As the first decoder, This is the second decoder.
5. The annotation method according to claim 4, characterized in that, The reconstruction loss functions of the first decoder and the second decoder are: ; n represents the number of cells input for single-cell ribonucleic acid data, and m represents the number of cells input for spatial transcriptome data. For single-cell RNA data, the input gene expression of the i-th cell is... This represents gene expression after reconstruction of the i-th cell from single-cell ribonucleic acid data. For the input gene expression of the j-th cell in the spatial transcriptome data, This represents the gene expression in the j-th cell after reconstruction of the spatial transcriptome data.
6. The annotation method according to claim 1, characterized in that, The steps of annotating the spatial transcriptome data by cell type based on a preset distance function combined with the first latent feature and the second latent feature include: The similarity between the first latent feature and the second latent feature is measured based on a preset distance function, and a mapping relationship is constructed. Cell type annotation is performed on the spatial transcriptome data using the mapping relationship.
7. The annotation method according to claim 1, characterized in that, The step of aligning the distributions of the first latent feature and the second latent feature in a common space includes: The first and second latent features are fed into a domain classifier for adversarial learning; wherein, the domain alignment loss function in adversarial learning is the binary cross-entropy. ; ,in This represents the true domain label of sample i. The domain classifier predicts the probability that sample i belongs to the source domain.
8. A cell type annotation system, characterized in that, The annotation system includes: The extraction module is used to identify common genes in single-cell data and spatial transcriptome data, and to extract the expression of the first gene corresponding to the common gene in the single-cell data and the expression of the second gene corresponding to the common gene in the spatial greening group data. The extraction module is further configured to use a common encoder to extract features from the expression of the first gene to obtain a first latent feature, and to extract features from the expression of the second gene to obtain a second latent feature; An alignment module is used to align the distributions of the first latent feature and the second latent feature in a common space; The annotation module annotates the spatial transcriptome data by cell type based on a preset distance function combined with the first latent feature and the second latent feature; The steps of extracting features from the expression of the first gene using a common encoder to obtain a first latent feature, and extracting features from the expression of the second gene to obtain a second latent feature, include: The first gene expression and the second gene expression are input into a common encoder to obtain the first latent feature and the second latent feature; Wherein, XSC represents the first gene expression, XST represents the second gene expression, ZSC represents the first latent trait, and ZST represents the second latent trait; ZSC=MLP1 (Attention(MLP2 (XSC))); ZST = MLP1 (Attention(MLP2 (XST))), where MLP1 and MLP2 are multilayer perceptrons, and Attention is an attention mechanism; The common encoder consists of a multilayer perceptron and an attention mechanism. Through the hierarchical computation of the encoder, the first layer MLP performs preliminary feature transformation on the gene expression data. Then, the attention mechanism focuses on the key gene features related to cell type and suppresses irrelevant noise. Finally, the second layer MLP maps the high-dimensional gene expression into low-dimensional first and second latent features, realizing the feature representation of the two types of data in the common semantic space. Prior to the step of aligning the distributions of the first latent feature and the second latent feature in a common space, the method includes: The first latent feature is input into the classifier, and supervised learning is performed using known cell type labels to optimize the classification accuracy of the cell type based on the first latent feature.