Automatic cell type annotation method for single-cell chromatin open data
By integrating TF-IDF preprocessing and MLP-GCN, the problems of insufficient stability and recognition ability in the automatic annotation methods of single-cell chromatin openness data were solved, and efficient and accurate cell type annotation was achieved, especially with excellent performance on rare cells and unknown cell types.
Patent Information
- Application Number
- CN202510778939.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-06-11
- Publication Date
- 2025-09-16
AI Technical Summary
Existing technologies for automatic annotation of single-cell chromatin open data have problems such as poor stability of heterogeneous datasets, failure to mine complex similarities between cells, lack of ability to identify unknown cell types, and lack of interpretability. They perform particularly poorly on rare cell types and cross-tissue and cross-sequencing technology datasets.
By adopting a method integrating TF-IDF-based data preprocessing, training set type balancing, multi-layer perceptron (MLP) training, low-dimensional representation and graph convolutional network (GCN), automatic annotation of single-cell chromatin accessibility data is achieved through comparative learning and multi-level annotation.
It improves the annotation accuracy of rare cell types, enhances the ability to identify unknown cell types, and improves the stability and interpretability of heterogeneous datasets, significantly outperforming existing methods, especially in performance on cross-batch, cross-tissue, and cross-species datasets.
Smart Images
Figure CN120656560A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of bioinformatics, and in particular to a method for automatic annotation of cell types based on single-cell chromatin open data. Background Art
[0002] Single-cell sequencing technology has made significant progress, significantly deepening our understanding of cellular heterogeneity in a variety of biological systems. Single-cell chromatin accessibility sequencing (scCAS) is a technique that captures the chromatin landscape that regulates transcription by analyzing chromatin accessibility. Its rapid development has provided new perspectives for studying epigenomic heterogeneity and the complex mechanisms of gene regulation at single-cell resolution.
[0003] Many downstream analyses of scCAS data (such as differential open peak analysis, cell type-specific motif analysis, transcription factor activity inference, etc.) usually require accurate cell type annotations in advance. The traditional mainstream annotation method is to cluster cells and then manually annotate them according to the peaks related to marker genes. As high-throughput sequencing technology drives the continuous expansion of scCAS data, traditional manual annotation is not only time-consuming and labor-intensive, but also poses computational feasibility challenges for large-scale data sets. These difficulties highlight the urgent need for automated cell annotation methods - methods that can efficiently use annotated data sets to accurately annotate newly generated data.
[0004] Although many automatic annotation methods suitable for single-cell transcriptome data have been developed, the effectiveness of these methods in scCAS data is significantly reduced due to the inherent high noise and extreme sparsity of scCAS data. To address the unique challenges of scCAS data, researchers have recently developed several dedicated methods. For example, Chen et al. proposed the first scCAS-specific method EpiAnno, which embeds cells into a latent space for annotation through a Bayesian neural network; SANGO developed by Zeng et al. integrates DNA sequence information and has become the most advanced scCAS cell annotation method. In addition, traditional machine learning methods such as support vector machines (SVM), random forests (RF) and K-nearest neighbors (KNN) have also shown excellent and robust performance in cell annotation.
[0005] However, existing methods still have major limitations:
[0006] 1. Poor stability of heterogeneous datasets. With the surge in scCAS datasets across tissues and sequencing technologies, existing methods have difficulty maintaining stable performance in heterogeneous datasets, especially for the annotation of rare cell types.
[0007] 2. Failure to mine complex similarities between cells. scCAS data often contain complex similarities between cells, and these features are crucial for capturing subtle intercellular relationships and accurate annotation.
[0008] 3. Lack of ability to identify unknown cell types. When unknown cell types appear in the test set that are not covered in the training set (such discoveries can reveal unknown cell differentiation and lineage relationships), existing scCAS-specific cell type annotation methods lack the ability to identify new cells.
[0009] 4. Lack of interpretability. Most methods lack interpretability, hindering researchers from analyzing cellular functions and underlying biological mechanisms from annotation results. Summary of the Invention
[0010] In view of the deficiencies of the existing technology, the technical problem to be solved by the present invention is to provide a method for automatic annotation of cell types for single-cell chromatin open data.
[0011] The technical solution of the present invention to solve the above technical problem is to provide a method for automatic annotation of cell types for single-cell chromatin open data, characterized in that the method comprises the following steps:
[0012] Step 1: Data preprocessing based on TF-IDF: First, the given cell type label C train Encoded as data label y train and label the cell type C train The original scCAS training set and the one without cell type labels C train The scCAS original set to be annotated is combined into the scCAS count matrix n0 represents the total number of sample cells after merging, p0 represents the initial peak number, and then the noise is reduced to obtain the filtered matrix p represents the number of peaks after filtering; then the peaks are weighted using TF-IDF conversion and normalized to obtain the preprocessed data set
[0013] Step 2: Type balancing of the training set: The preprocessed dataset X obtained in step 1 is processed Split to get the training set and the set to be annotated n train Represents the total number of sample cells in the training set, n target Represents the total number of sample cells in the set to be annotated; then the training set type balancing technology is used to process the training set to obtain the processed training set n train_processed Represents the total number of sample cells in the training set after processing;
[0014] Step 3: Use the processed training set X obtained in step 2 train_processedThe MLP is trained through contrastive learning to obtain the trained MLP;
[0015] Step 4: Multi-level annotation using low-dimensional representation: Use the trained MLP to obtain low-dimensional representations of cells, and then obtain the first soft label annotation matrix based on the low-dimensional representations of each cell. Second soft label annotation matrix and cell type recognition results; then according to the second soft label annotation matrix Calculate cell type specific chromatin accessible peaks and annotate the first soft label matrix Second soft label annotation matrix Integrate with the cell type recognition results to obtain the final annotation result C target .
[0016] Compared with the prior art, the present invention has the following beneficial effects:
[0017] (1) Accurate Annotation Performance: We compared the annotation performance of our method with that of the baseline model within the data. We randomly divided the cells in each dataset into five folds, trained the model using four folds of data at a time, and performed cell type annotation on the remaining fold. Our method achieved near-optimal performance on all six datasets, as confirmed by a one-sided paired Wilcoxon signed-rank test, which demonstrated that our method significantly outperformed the baseline method on all metrics.
[0018] (2) Rare cell type annotation: This paper further conducted a special evaluation of the rare cell types with the smallest proportion in the six datasets (using rare cells as positive samples and the remaining cells as negative samples) and compared them with SANGO. Although the overall performance of this paper was comparable to SANGO in some datasets (such as Heart), it showed a significant advantage in rare cell type annotation.
[0019] (3) Unknown cell type identification: To simulate the presence of unknown cell types in the test set, the present invention first divides the dataset into a training set (80%) and a test set (20%). The present invention then removes specific cell types from the training set, leaving them only in the test set. When different cell types are considered unknown, the present invention demonstrates high recall rates across all experimental datasets, demonstrating that it can successfully identify all cells belonging to the unknown cell type.
[0020] (4) Interpretability: To verify whether the cell type-specific peaks identified in this study can reveal biological mechanisms, we systematically analyzed the cell type-specific peaks in the ThymusA dataset (which contains four cell types closely related to immune system diseases: thymocytes, vascular endothelial cells, antigen-presenting cells, and thymic epithelial cells). Based on multiple downstream analyses, we found that this study can provide important clues for the study of genetic and cellular mechanisms, deepening our understanding of the complex biological processes of cell type specificity.
[0021] (5) Annotation performance of heterogeneous datasets: The present invention further constructed cross-batch datasets, cross-tissue datasets, and cross-species datasets for experiments. In multiple tasks, the present invention showed better performance, fully demonstrating its flexibility in real-world cell annotation tasks.
[0022] (6) Running time: The annotation time of the proposed method for datasets of different sizes is significantly shorter than that of EpiAnno and SANGO. BRIEF DESCRIPTION OF THE DRAWINGS
[0023] Figure 1 It is the overall flow chart of the present invention;
[0024] Figure 2 The results of the five-fold cross-validation comparison experiment on the dataset of Example 1 of the present invention are as follows;
[0025] Figure 3 This is the cross-batch comparison experiment result of Example 1 of the present invention;
[0026] Figure 4 These are the results of cross-tissue and cross-species comparative experiments in Example 1 of the present invention;
[0027] Figure 5 This is the comparative experimental result of unknown cell type identification in Example 1 of the present invention. DETAILED DESCRIPTION
[0028] The specific embodiments of the present invention are given below. The specific embodiments are only used to further illustrate the present invention and do not limit the scope of protection of the present invention.
[0029] The present invention provides a method for automatic annotation of cell types based on single-cell chromatin openness data (hereinafter referred to as the method), characterized in that the method comprises the following steps:
[0030] Step 1: Data preprocessing based on TF-IDF: First, the given cell type label C train Encoded as data label y train and label the cell type C train The original scCAS training set and the one without cell type labels C trainThe scCAS original set to be annotated is combined into the scCAS count matrix n0 represents the total number of sample cells after merging, p0 represents the initial peak number (cell chromatin accessible peak), and then the noise is reduced to obtain the filtered matrix p represents the number of peaks after filtering; then the peaks are weighted using TF-IDF (term frequency-inverse document frequency) conversion and normalized to obtain the preprocessed data set.
[0031] Preferably, in step 1, noise reduction is achieved by filtering the count matrix X raw Open peaks in less than 1% of cells.
[0032] Preferably, in step 1, the specific calculation process of TF-IDF conversion is as follows:
[0033]
[0034] Preferably, in step 1, normalization is specifically:
[0035]
[0036] Step 2: Type balancing of the training set: The preprocessed dataset X obtained in step 1 is processed Split to get the training set and the set to be annotated n train Represents the total number of sample cells in the training set, n target Represents the total number of sample cells in the set to be annotated; Then, in response to the common type imbalance problem in the scCAS dataset, the training set type balancing technology is used to process the training set to obtain the processed training set n train_processed Represents the total number of sample cells in the training set after processing;
[0037] Preferably, in step 2, the training set type balancing technique is specifically as follows: first, a threshold K is set; then, for cell types whose cell number exceeds K, downsampling is performed to reduce the cell number to K; for cell types whose cell number is less than K, an oversampling technique based on a mask strategy is used to generate synthetic cells to supplement K; and then the processed training set X is obtained. train_processed .
[0038] Preferably, in step 2, the oversampling technique is: randomly setting 15% of non-zero elements to zero.
[0039] Step 3: Use the processed training set X obtained in step 2 train_processedThe MLP (Multi-layer Perceptron) is trained through contrastive learning to obtain a trained MLP, which can effectively learn the low-dimensional representation of each cell;
[0040] Preferably, in step 3, contrastive learning is a representation learning method based on similarity measurement. By constructing positive and negative sample pairs, the contrastive learning method can learn the low-dimensional representation of data. In the contrastive learning framework, the positive and negative sample pair construction method is used: for the processed training set X train_processed The peak data x of the i-th cell in i , randomly select the peak data of a cell with the same cell type as the i-th cell Construct the positive sample pair of the i-th cell At the same time, randomly select the peak data of a cell whose cell type is different from the i-th cell type Construct the negative sample pair of the i-th cell For the processed training set X train_processed Each cell in the positive and negative sample pair construction method is applied, and finally n train positive sample pairs and n train negative sample pairs.
[0041] Preferably, in step 3, the loss function Loss of the MLP based on positive and negative sample pairs is:
[0042]
[0043] In formula (3), represents the low-dimensional representation of the i-th cell, represents the low-dimensional representation of the cell in the positive sample pair of the i-th cell, represents the low-dimensional representation of the cell in the negative sample pair of the i-th cell; cos(·) represents the cosine similarity, which is calculated as follows:
[0044]
[0045] Preferably, in step 3, in the low-dimensional space, cells of the same type are clustered together, while cells of different types are clearly separated.
[0046] Step 4: Multi-level annotation using low-dimensional representation: Use the trained MLP to obtain low-dimensional representations of cells, and then obtain the first soft label annotation matrix based on the low-dimensional representations of each cell. Second soft label annotation matrix and cell type recognition results; then according to the second soft label annotation matrix Calculate cell type specific chromatin accessible peaks and annotate the first soft label matrix Second soft label annotation matrix Integrate with the cell type recognition results to obtain the final annotation result C target ; Cell type-specific chromatin accessible peaks and final annotation results C target as the final output.
[0047] Preferably, in step 4, the first soft label annotation matrix is obtained The specific steps are as follows:
[0048] A1. The processed training set X obtained in step 2 tarain_processed and the set to be annotated X target Input into the trained MLP to obtain the low-dimensional representation of the processed training set and a low-dimensional representation of the set to be annotated d represents the dimension of the low-dimensional representation;
[0049] A2. Low-dimensional representation of the processed training set Calculate the algebraic centroid of the k-th cell And stack K centroid vectors into a centroid matrix
[0050]
[0051] In formulas (5) and (6), Represents the low-dimensional representation of the processed training set The index set of cells belonging to the kth type in the training set, K represents the number of cell types in the training set;
[0052] A3. Low-dimensional representation of the set to be annotated For each test cell in the , calculate the centroid matrix of each test cell The cosine similarity of each centroid in:
[0053]
[0054] Then perform normalization:
[0055]
[0056] Then Assumed as the probability that the i-th test cell belongs to the k-th class; then use all the probabilities Generate a cosine similarity matrix and use it as the first soft label annotation matrix
[0057] Preferably, in step 4, a second soft label annotation matrix is obtained The specific steps are as follows:
[0058] B1. The training set X obtained in step 2 trainInput into the trained MLP to obtain a low-dimensional representation of the training set Low-dimensional representation of the training set Specific cells in Apply the K nearest neighbor graph construction strategy to obtain the training set X train The adjacency matrix A train ;
[0059] Preferably, in step B1, the low-dimensional representation of the training set Specific cells in Find the nearest m cells based on cosine similarity, and these m cells are considered as cells Neighbors of Redefine the training set X train The adjacency matrix A train , the adjacency matrix A train Elements in for:
[0060]
[0061] At the same time, the union of the low-dimensional representation of the training set and the low-dimensional representation of the set to be annotated is used Apply the same K-nearest neighbor graph construction strategy to obtain another adjacency matrix
[0062] B2. Add the adjacency matrix A train and the adjacency matrix A all Add them together and binarize the result to get the final adjacency matrix
[0063] B3. The final adjacency matrix A final , the training set X obtained in step 2 train , the set to be annotated X obtained in step 2 target And the data label y obtained in step 1 train Input to Graph Convolutional Networks (GCN) for semi-supervised learning to obtain the second soft label annotation matrix
[0064] Preferably, in step 4, the specific steps of obtaining the cell type identification result are as follows:
[0065] C1, the training set X obtained in step 2 tarain Input into the trained MLP to obtain a low-dimensional representation of the training set The low-dimensional representation of the training set According to K different known cell types, divided into K different training subsets
[0066] C2. Use each training subset separately Calculate the convex hull S of each cell type in low-dimensional space k :
[0067]
[0068] In formula (10), λ j is the coefficient of the point convex combination, x k，j represents the low-dimensional representation of the jth cell in the space associated with cell type k;
[0069] C3. Treating low-dimensional representation of the annotation set For each cell in , determine whether it is located in the convex hull of the known cell type; if it is, it is regarded as a cell of the known cell type; if not, it is regarded as a cell of the unknown cell type, and then the cell type recognition results of all cells are obtained.
[0070] Preferably, in step 4, obtaining cell type specific chromatin accessible peaks is: using a second soft label annotation matrix Calculate cell type-specific chromatin accessible peaks based on mutual information (MI). The specific steps are as follows:
[0071] Construct a feature selector F∈{0,1} for GCN nodes p , and define the feature subset X F , the feature selector filters relevant features as follows:
[0072]
[0073] In formula (11), represents the features in the i-th cell that are not masked by F;
[0074] Then optimize the feature selector F by maximizing the mutual information to obtain the optimized feature selector F t , the specific optimization method is:
[0075]
[0076] In formula (12), the calculation of conditional entropy is defined as:
[0077]
[0078] Then the set to be annotated X obtained in step 2 target According to K different known cell types, divided into K different subsets to be annotated Then, for each known cell type, the corresponding subset to be annotated Apply the optimized feature selector F t , and obtain the cell type-specific peak selection results; then use the Sigmoid function to normalize the cell type-specific peak selection results to obtain the importance score of the cell type-specific peak, and select the top K peaks with the highest importance as the cell type-specific chromatin accessible peaks.
[0079] Preferably, in step 4, the final annotation result C is obtained target Specifically: the first soft label annotation matrix and the second soft label annotation matrix Add together to get the final soft label annotation matrix For the set X to be annotated target Each cell i in is annotated by the final soft label matrix The i-th row vector in The maximum value in determines the final label of cell i. The cell type corresponding to the maximum value is the most likely type of cell i. The most likely types of all cells are used to generate a vector as the output result y target ∈R K , and then output the result y target After decoding, the initial annotation result C is obtained target_raw ; Then in the initial annotation result C target_raw In the process, according to the cell type recognition results, the annotation results of cells of unknown cell types are replaced with unknown to obtain the final annotation result C target .
[0080] Example 1:
[0081] The present invention was compared with six baseline cell type annotation methods, including two methods designed specifically for scCAS data and four traditional machine learning methods.
[0082] The two scCAS-specific methods are:
[0083] SANGO: SANGO encodes the peak genomic sequence into a low-dimensional embedding, then iteratively reconstructs the cell's peak statistics through a fully connected network, and annotates the query cell based on a graph transformer network. Follow the GitHub tutorial (https: / / github.com / biomed-AI / SANGO#Tutorial) and run it with default parameters.
[0084] EpiAnno: EpiAnno is a probabilistic generative model combined with a Bayesian neural network for automatically annotating scCAS data in a supervised manner. Experiments were performed using the default parameters following the steps provided in the GitHub code (https: / / github.com / xychen16 / EpiAnno / blob / master / code / demo.ipynb).
[0085] Four traditional machine learning approaches include:
[0086] Support Vector Machine (SVM): SVM is a supervised learning algorithm that classifies data by finding an optimal hyperplane. Based on recent benchmark research, we used the sklearn.svm module from scikit-learn and selected the linear kernel function.
[0087] Random Forest (RF): RF is an ensemble learning method that constructs multiple decision trees and outputs the mode of their classification results. This experiment uses the sklearn.ensemble.RandomForestClassifier class from scikit-learn, with default parameters.
[0088] K-nearest neighbor (KNN): KNN is a simple instance-based algorithm that classifies based on the majority label of the nearest neighbors in feature space. Use the scikit-learn sklearn.neighbors.KNeighborsClassifier class with n_neighbors = 9 and n_neighbors = 50 for the KNN9 and KNN50 models, respectively.
[0089] To evaluate the model's ability to identify unknown cell types, a support vector machine (SVM) with a rejection option was used. This algorithm has been shown to accurately identify unknown cell types in recent benchmark studies. Following the recommendations of the benchmark study, a posterior probability threshold of 0.7 was set, and cells below this threshold were labeled as novel.
[0090] In this example, 11 scCAS datasets were systematically collected, covering different species, platforms, data scales, dimensions, type imbalance levels, and major cell type proportions for comprehensive benchmark evaluation.
[0091] Human tissue datasets: Two sets of cellular data from human tissues were collected: a melanoma dataset (Melanoma) containing 598 cells derived from a SOX10 knockout melanoma cell line obtained from short-term culture samples of two patients; and spleen and thymus datasets (SpleenA and ThymusA), generated using sci-ATAC-seq3 technology and derived from human fetal samples. These two datasets have higher cell counts and spike counts, enabling a more comprehensive assessment of annotation performance in complex datasets.
[0092] Mouse tissue datasets: To explore the applicability of the method across diverse species and data sources, six datasets (ThymusB, Liver, Heart, LungA, LungB, and SpleenB) were collected from the mouse sci-ATAC-seq atlas. These datasets, generated using a combined index detection technology (sci-ATAC-seq), contain varying degrees of imbalance in cell numbers and types, providing a diverse testing environment for evaluation.
[0093] Cross-species annotation experiment datasets: To conduct cross-species cell type annotation experiments, additional human and mouse brain tissue datasets were collected. The human brain dataset primarily consists of 130,418 cells, obtained from postmortem brain tissue of Alzheimer's patients and cognitively healthy controls, covering all major brain cell types. The mouse brain dataset, derived from the mouse sci-ATAC-seq atlas, includes cerebellar cells from 8-week-old mice. To facilitate cross-species comparisons, the excitatory and inhibitory neuronal subtypes in the mouse data were merged (as these undefined subtypes cannot be directly mapped to the human data).
[0094] The specific comparative experiments mainly include five-fold cross-validation experiments of the dataset, cross-batch experiments using batch labels within the dataset, cell annotation experiments of cross-tissue and cross-species data, and unknown cell type identification experiments.
[0095] Hyperparameters set when implementing the present invention:
[0096] An MLP is used to learn the low-dimensional representations. This MLP consists of two fully connected layers. The input layer receives a cell feature vector whose dimensions correspond to the dimensionality of the input features. The first fully connected layer projects the input into a latent space with 144 neurons. This is then processed using a sigmoid activation function and L2 normalization to ensure stability and better representation learning. The second layer, another fully connected layer, further transforms the latent representation into a 64-dimensional output space, corresponding to the learned low-dimensional embedding of the cell.
[0097] The cosine similarity function is used to calculate the similarity between the cell embeddings in the sample pairs. This similarity serves as the basic metric for contrastive learning, enabling the model to effectively distinguish between similar and dissimilar cell states. The model is optimized using the Adam optimizer with an initial learning rate of 0.001 and a learning rate scheduler that reduces the learning rate to 0.1 times the original rate after 30 epochs to improve convergence. The training process adopts a small batch strategy, with the batch size set to 10% of the size of the positive sample pair dataset. In addition, an early stopping mechanism based on the loss value is implemented: if the loss value is lower than 0.05, or the average loss change over 5 consecutive epochs is less than 0.0001, training is terminated to prevent overfitting and ensure computational efficiency.
[0098] GCN was used in the graph learning phase. This GCN model consists of three graph convolutional layers. The first GCNConv layer converts the input features into a 128-dimensional latent space. The second GCNConv layer further reduces the feature dimensionality to 64 dimensions. The third GCNConv layer finally outputs a dimension corresponding to the number of cell types. Each graph convolution layer uses the ReLU activation function for nonlinear transformation, and the final output is processed by the softmax function to obtain the classification probability. The model is optimized using the Adam optimizer with an initial learning rate of 0.0001.
[0099] from Figure 2 It can be seen that the present invention achieves the best performance on all six data sets. After performing a one-sided paired Wilcoxon signed rank test on these indicators, it is proved that the present invention significantly outperforms the baseline method in all indicators.
[0100] from Figure 3 、 4 As can be seen in the figure, in the annotation task of heterogeneous datasets such as cross-batch, cross-tissue and cross-species, the present invention can still show certain advantages in all indicators.
[0101] from Figure 5 As can be seen, in the unknown cell type recognition experiment, the present invention demonstrated high recall rates across all datasets, indicating that it was able to successfully identify all cells belonging to the unknown cell type. In terms of the F1 score, which comprehensively considers recall and precision, the present invention consistently demonstrated superior overall performance.
[0102] Any matters not described in the present invention are applicable to the prior art.
Claims
1. A method for automatic annotation of cell types based on single-cell chromatin openness data, characterized in that: The method comprises the following steps: Step 1: Data preprocessing based on TF-IDF: First, the given cell type label C train Encoded as data label y train and label the cell type C train The original scCAS training set and the one without cell type labels C train The scCAS original set to be annotated is combined into the scCAS count matrix n0 represents the total number of sample cells after merging, p0 represents the initial peak number, and then the noise is reduced to obtain the filtered matrix p represents the number of peaks after filtering; then the peaks are weighted using TF-IDF conversion and normalized to obtain the preprocessed data set Step 2: Type balancing of the training set: The preprocessed dataset X obtained in step 1 is processed Split to get the training set and the set to be annotated n train Represents the total number of sample cells in the training set, n target Represents the total number of sample cells in the set to be annotated; then the training set type balancing technology is used to process the training set to obtain the processed training set n train_processed Represents the total number of sample cells in the training set after processing; Step 3: Use the processed training set X obtained in step 2 train_processed The MLP is trained through contrastive learning to obtain the trained MLP; Step 4: Multi-level annotation using low-dimensional representation: Use the trained MLP to obtain low-dimensional representations of cells, and then obtain the first soft label annotation matrix based on the low-dimensional representations of each cell. Second soft label annotation matrix and cell type recognition results; then according to the second soft label annotation matrix Calculate cell type specific chromatin accessible peaks and annotate the first soft label matrix Second soft label annotation matrix Integrate with the cell type recognition results to obtain the final annotation result C target .
2. The method according to claim 1, wherein: In step 1, noise reduction is achieved by filtering the count matrix X raw less than 1% of cells have open peaks; In step 1, the specific calculation process of TF-IDF conversion is as follows: In step 1, normalization is specifically:
3. The method according to claim 1, wherein: In step 2, the training set type balancing technique is as follows: first set a threshold K; then, for cell types with more than K cells, downsample to reduce the number of cells to K; for cell types with less than K cells, use mask-based oversampling to generate synthetic cells to supplement K; and then obtain the processed training set X train_processed ; In step 2, the oversampling technique is to randomly set 15% of the non-zero elements to zero.
4. The method according to claim 1, wherein: In step 3, in the contrastive learning framework, the positive and negative sample pair construction method is used: for the processed training set X train_processed The peak data x of the i-th cell in i , randomly select the peak data of a cell with the same cell type as the i-th cell Construct the positive sample pair of the i-th cell At the same time, randomly select the peak data of a cell whose cell type is different from the i-th cell type Construct the negative sample pair of the i-th cell For the processed training set X train_processed Each cell in the positive and negative sample pair construction method is applied, and finally n train positive sample pairs and n train negative sample pairs; In step 3, the loss function Loss of the MLP based on positive and negative sample pairs is: In formula (3), represents the low-dimensional representation of the i-th cell, represents the low-dimensional representation of the cell in the positive sample pair of the i-th cell, represents the low-dimensional representation of the cell in the negative sample pair of the i-th cell; cos(·) represents the cosine similarity, which is calculated as follows:
5. The method according to claim 1, wherein: In step 4, the first soft label annotation matrix is obtained The specific steps are as follows: A1. The processed training set X obtained in step 2 train_prpcessed and the set to be annotated X target Input into the trained MLP to obtain the low-dimensional representation of the processed training set and a low-dimensional representation of the set to be annotated d represents the dimension of the low-dimensional representation; A2. Low-dimensional representation of the processed training set Calculate the algebraic centroid of the k-th cell And stack K centroid vectors into a centroid matrix In formulas (5) and (6), Represents the low-dimensional representation of the processed training set The index set of cells belonging to the kth type in the training set, K represents the number of cell types in the training set; A3. Low-dimensional representation of the set to be annotated For each test cell in the , calculate the centroid matrix of each test cell The cosine similarity of each centroid in: Then perform normalization: Then Assume that the probability that the i-th test cell belongs to the k-th class; then use all the probabilities Generate a cosine similarity matrix and use it as the first soft label annotation matrix 6. The method according to claim 1, wherein: In step 4, the second soft label annotation matrix is obtained The specific steps are as follows: B1. The training set X obtained in step 2 train Input into the trained MLP to obtain a low-dimensional representation of the training set Low-dimensional representation of the training set Specific cells in Apply the K nearest neighbor graph construction strategy to obtain the training set X train The adjacency matrix A train ; At the same time, the union of the low-dimensional representation of the training set and the low-dimensional representation of the set to be annotated is used Apply the same K-nearest neighbor graph construction strategy to obtain another adjacency matrix B2. Add the adjacency matrix A train and the adjacency matrix A all Add them together and binarize the result to get the final adjacency matrix B3. The final adjacency matrix A final , the training set X obtained in step 2 train , the set to be annotated X obtained in step 2 target And the data label y obtained in step 1 train Input into the graph convolutional network for semi-supervised learning to obtain the second soft label annotation matrix 7. The method according to claim 6, wherein: In step B1, the low-dimensional representation of the training set Specific cells in Find the nearest m cells based on cosine similarity, and these m cells are considered as cells Neighbors of Redefine the training set X train The adjacency matrix A train , the adjacency matrix A train Elements in for:
8. The method according to claim 1, wherein: In step 4, the specific steps for obtaining the cell type identification result are as follows: C1, the training set X obtained in step 2 train Input into the trained MLP to obtain a low-dimensional representation of the training set The low-dimensional representation of the training set According to K different known cell types, divided into K different training subsets C2. Use each training subset separately Calculate the convex hull S of each cell type in low-dimensional space k : In formula (10), λ j is the coefficient of the point convex combination, x k,j represents the low-dimensional representation of the jth cell in the space associated with cell type k; C3. Treating low-dimensional representations of annotation sets For each cell in , determine whether it is located in the convex hull of the known cell type; if it is, it is considered a cell of the known cell type; If it is not located, it is regarded as a cell of unknown cell type, and the cell type identification results of all cells are obtained.
9. The method according to claim 1, wherein: In step 4, the specific steps to obtain cell type-specific chromatin accessible peaks are as follows: Construct a feature selector F∈{0,1} for GCN nodes p , and define the feature subset X F , the feature selector filters relevant features as follows: In formula (11), represents the features in the i-th cell that are not masked by F; Then optimize the feature selector F by maximizing the mutual information to obtain the optimized feature selector F t , the specific optimization method is: max F MI(and target ,X F )=H(y target )-H(and target |X=X F ) (12) In formula (12), the calculation of conditional entropy is defined as: Then the set to be annotated X obtained in step 2 target According to K different known cell types, divided into K different subsets to be annotated Then, for each known cell type, the corresponding subset to be annotated Apply the optimized feature selector F t , and obtain the cell type-specific peak selection results; then use the Sigmoid function to normalize the cell type-specific peak selection results to obtain the importance score of the cell type-specific peak, and select the top K peaks with the highest importance as the cell type-specific chromatin accessible peaks.
10. The method according to claim 1, wherein: In step 4, the final annotation result C is obtained target Specifically: the first soft label annotation matrix and the second soft label annotation matrix Add together to get the final soft label annotation matrix For the set X to be annotated target Each cell i in is annotated by the final soft label matrix The i-th row vector in The maximum value in determines the final label of cell i. The cell type corresponding to the maximum value is the most likely type of cell i. The most likely types of all cells are used to generate a vector as the output result y target ∈R K , and then output the result t target After decoding, the initial annotation result C is obtained target_raw ; Then in the initial annotation result C target_raw In the process, according to the cell type recognition results, the annotation results of cells of unknown cell types are replaced with unknown to obtain the final annotation result C target .