A cell class determination method and system

Abstract

The application discloses a cell category determination method and system, and relates to the technical field of cell classification. The method comprises the following steps: constructing an scDRLN model; processing a source domain data set by using a self-step learning mechanism, and determining source domain training data sets and cell categories of corresponding single-cell RNA sequencing data of the source domain training data sets under each iteration number of a small-batch gradient descent method; taking the minimum value of a total loss function as a target, and training the scDRLN model according to the source domain training data sets under each iteration number of the small-batch gradient descent method, the cell categories of the corresponding single-cell RNA sequencing data of the source domain training data sets and a target domain data set, and the first branch of the trained scDRLN model is used for determining a cell category. The application can fully utilize the distribution and structure information of data to improve the accuracy of cell category recognition results, and can recognize rare category cells.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of cell classification, and in particular to a cell class determination method and system. BACKGROUND

[0002] Single-cell RNA sequencing technology (scRNA-seq) can perform high-resolution analysis of gene expression, reveal the heterogeneity of cell populations, and gain in-depth understanding of the behavior and mechanisms of individual cells. Compared with early batch RNA-seq data, scRNA-seq data is more sparse and has a higher dropout rate. In scRNA-seq analysis, it is often necessary to integrate data from multiple batches. Due to factors such as sequencing platforms or experimental operations, joint analysis of multiple batches of data often has batch effects in practice, which will have a negative impact on downstream analysis, such as cell class recognition. For example, distribution differences between batches can affect the intrinsic differences between cell classes, leading to incorrect labeling of cell classes. Therefore, there is an urgent need to develop an automated method for cell class labeling and to correct batch effects.

[0003] Many machine learning-based methods have been used to identify cell classes from scRNA-seq data. Single-cell consensus clustering (SC3) calculates a consensus matrix by averaging multiple similarity matrices, and uses the k-means algorithm for clustering. SIMLR clusters cells into different clusters according to the similarity of cells, and then manually assigns each cluster a cell class label for cell class recognition. However, the above methods cannot be directly transferred to related data sets, and require labor-intensive repetitive work. With the development of sequencing technology, classification algorithms based on well-annotated public data sets have been developed for cell class recognition. For example, SingleR calculates the similarity between each cell in the test set and each cell in the known reference set, and assigns labels to the test set according to the similarity. In addition, scmap assigns cell class labels by querying similar cells in the reference database. To explore the non-linear relationship between genes, scPred performs singular value decomposition on the gene expression matrix to obtain limited information features and train a support vector machine model. However, the premise of the machine learning algorithm is that the training set and test set data come from the same distribution, but single-cell data sets may have distribution mismatches due to batch effects and other factors, which will affect the performance of the model.

[0004] In order to solve the problem of data distribution mismatch between the training set and the test set due to batch effects and other factors, MNN uses a mutual neighbor matching algorithm to find the nearest neighbors and shared "anchor" cells of each batch, and corrects the gene expression of all cells in each batch based on these anchors, thereby reducing the batch effect. LIGER integrates multiple single-cell datasets and uses integrated nonlinear matrix factorization (iNMF) to alleviate the batch effect. Harmony uses principal component analysis (PCA) to embed scRNA data into a low-dimensional space and uses fuzzy clustering and iterative algorithms to alleviate the batch effect. DESC is an unsupervised deep embedding algorithm that clusters single-cell RNA data by iteratively optimizing the clustering objective function to correct the batch effect. Seurat uses CCA to improve batch effects and uses anchors to annotate cell categories, but does not fully utilize the well-labeled cell category information in the reference dataset.

[0005] In addition, some transfer learning methods have been used for batch correction and cell category recognition. BERMUDA detects cell clusters from different batches through clustering, uses Spearman to identify similar clusters between batches, and is used to train an autoencoder with MMD loss for batch correction. A reasonable assumption is that batch correction and cell classification are interdependent. Specifically, effective correction has a conductive effect on classification, while correct label information is beneficial to the correction process. It is worth noting that the above methods can make cell distributions consistent between different datasets, but cannot annotate cell categories. ItClust is a supervised clustering algorithm based on transfer learning that uses the distance between cluster centers and cell embeddings to predict cell categories, but this pre-training model has a catastrophic forgetting problem, which may lead to information loss and ignore the batch effect between the reference dataset and the target domain dataset.

[0006] In summary, batch effects in single-cell RNA sequencing (scRNA-seq) data remain a challenge for sample integration. Existing batch correction methods fail to fully utilize the distribution and structural information of the data, resulting in inaccurate final identification of cell categories. At the same time, existing cell category annotation methods fail to fully utilize well-annotated reference datasets, making it difficult to identify rare cell categories. Summary of the Invention

[0007] The purpose of the present invention is to provide a cell category determination method and system, which can fully utilize the distribution and structural information of data to improve the accuracy of cell category recognition results and can identify rare categories of cells.

[0008] To achieve the above object, the present invention provides the following solutions:

[0009] A method for determining a cell type, comprising:

[0010] Constructing a scDRLN model; the scDRLN model includes a first branch, a second branch, and a third branch; the first branch includes a feature extractor and a label classifier connected in sequence; the second branch includes a feature extractor and a domain discriminator connected in sequence; the third branch includes a feature extractor and a decoder connected in sequence;

[0011] Obtain a source domain dataset, a target domain dataset, and the cell categories of each single-cell RNA sequencing data corresponding to the source domain dataset; the source domain dataset includes n s samples; all samples in the source domain dataset are gene expression matrices processed from different batches of single-cell RNA sequencing data obtained from different tissues; the target domain dataset includes n t samples, all samples in the target domain dataset are gene expression matrices processed from single-cell RNA sequencing data of different batches obtained from the same tissue; the source domain dataset is different from the target domain dataset, n s and n t are all positive integers greater than 1, and n s and n t Not equal;

[0012] The source domain dataset is processed using a self-paced learning mechanism to determine a source domain training dataset for each iteration of a mini-batch gradient descent method;

[0013] Obtaining the cell category of each single-cell RNA sequencing data corresponding to the source domain training dataset at each iteration number of the mini-batch gradient descent method according to the source domain training dataset at each iteration number of the mini-batch gradient descent method and the cell category of each single-cell RNA sequencing data corresponding to the source domain dataset;

[0014] According to the value of the total loss function being minimum, the scDRLN model is trained by using the mini-batch gradient descent method according to the source domain training data set, the target domain data set and the cell category of each single cell RNA sequencing data corresponding to the source domain training data set of each iteration number of the mini-batch gradient descent method, to obtain a trained scDRLN model, and the first branch in the trained scDRLN model is used to determine the cell category; the total loss function includes a self-step regularization term, a cross-entropy loss function, a target domain data set reconstruction loss function, a global alignment loss function, a source domain data set alignment loss function and a semantic centroid alignment loss function; the cross-entropy loss function is determined based on the source domain training data set, the target domain data set and the first branch; the target domain data set reconstruction loss function is determined based on the target domain data set and the third branch; the global alignment loss function is determined based on the source domain training data set, the target domain data set and the second branch; the source domain data set alignment loss function is determined based on the source domain training data set and the feature encoder; the semantic centroid alignment loss function is determined based on the source domain training data set, the target domain data set, the label classifier and the feature encoder; and the self-step regularization term is determined according to the global alignment loss function.

[0015] A cell category determination system comprises:

[0016] A model construction module is configured to construct an scDRLN model, wherein the scDRLN model comprises a first branch, a second branch and a third branch; the first branch comprises a feature extractor and a label classifier connected in sequence; the second branch comprises a feature extractor and a domain discriminator connected in sequence; and the third branch comprises a feature extractor and a decoder connected in sequence.

[0017] An acquisition module is configured to acquire a source domain data set, a target domain data set and a cell category of each single cell RNA sequencing data corresponding to the source domain data set; the source domain data set comprises n s samples; all samples in the source domain data set are gene expression matrices processed from single cell RNA sequencing data of different batches obtained from different tissues; the target domain data set comprises n t samples, and all samples in the target domain data set are gene expression matrices processed from single cell RNA sequencing data of different batches obtained from the same tissue; the source domain data set is different from the target domain data set, n s and n t are positive integers greater than 1, and n s and n t are not equal.

[0018] A self-paced learning mechanism module, configured to process the source domain dataset using a self-paced learning mechanism to determine a source domain training dataset for each iteration of a small batch gradient descent method;

[0019] a training data cell category determination module, configured to obtain the cell category of each single-cell RNA sequencing data corresponding to the source domain training dataset at each iteration number of the mini-batch gradient descent method according to the source domain training dataset at each iteration number of the mini-batch gradient descent method and the cell category of each single-cell RNA sequencing data corresponding to the source domain dataset;

[0020] The training module is used to train the scDRLN model with the goal of minimizing the value of the total loss function, according to the source domain training dataset under each iteration of the mini-batch gradient descent method, the target domain dataset, and the cell category of each single-cell RNA sequencing data corresponding to the source domain training dataset under each iteration of the mini-batch gradient descent method, to obtain a trained scDRLN model, wherein the first branch in the trained scDRLN model is used to determine the cell category; the total loss function includes a self-paced regularization term, a cross entropy loss function, a target domain dataset reconstruction loss function, a global alignment loss function, a source domain dataset alignment loss function, and a semantic centroid alignment loss function. loss function; the cross entropy loss function is determined based on the source domain training dataset, the target domain dataset and the first branch; the target domain dataset reconstruction loss function is determined based on the target domain dataset and the third branch; the global alignment loss function is determined based on the source domain training dataset, the target domain dataset and the second branch; the source domain dataset alignment loss function is determined based on the source domain training dataset and the feature encoder; the semantic centroid alignment loss function is determined based on the source domain training dataset, the target domain dataset, the label classifier and the feature encoder; the self-paced regularization term is determined according to the global alignment loss function.

[0021] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0022] The present invention constructs a scDRLN model including a feature extractor, a label classifier, a domain discriminator and a decoder, and obtains a total loss function based on a cross entropy loss function, a target domain dataset reconstruction loss function, a global alignment loss function, a source domain dataset alignment loss function and a semantic centroid alignment loss function, and trains the scDRLN model with the goal of minimizing the total loss function; the cross entropy loss function is determined according to the feature extractor and the label classifier and is used to train the label classifier, and the target domain dataset reconstruction loss function is determined according to the feature extractor and the decoder and is used to reconstruct the input data, so that the feature extractor can better retain the characteristics of the target domain dataset. The global alignment loss function is determined according to the feature extractor and the domain discriminator, and is used to train the feature extractor and the domain discriminator, so that the domain discriminator cannot distinguish the batches, and solves the problem that the existing batch correction method fails to make full use of the distribution and structural information of the data, resulting in inaccurate cell categories finally identified. The source domain dataset alignment loss function is determined according to the feature encoder, and the feature encoder is trained. The semantic centroid alignment loss function is determined according to the feature encoder and the label classifier, and is used to train the feature encoder and the label classifier, so that the first branch finally obtained can make full use of the well-labeled reference dataset, and can recognize rare categories of cells. BRIEF DESCRIPTION OF THE DRAWINGS

[0023] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0024] Figure 1 A framework diagram of the cell type determination method provided by the present invention;

[0025] Figure 2 A flowchart of the preprocessing process provided by an embodiment of the present invention;

[0026] Figure 3 Flowchart of the cell type determination method provided by the present invention. DETAILED DESCRIPTION

[0027] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0028] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0029] An embodiment of the present invention provides a method for determining cell type, comprising:

[0030] Construct scDRLN model; Figure 1 As shown, the scDRLN model includes a first branch, a second branch and a third branch; the first branch includes a feature extractor and a label classifier connected in sequence; the second branch includes a feature extractor and a domain discriminator connected in sequence; the third branch includes a feature extractor and a decoder connected in sequence.

[0031] Obtain the source domain dataset, the target domain dataset, and the cell categories of each single-cell RNA sequencing data corresponding to the source domain dataset; the source domain dataset is a single-cell gene expression matrix obtained by sorting and processing the single-cell RNA sequencing data, including n s samples; all samples in the source domain dataset are gene expression matrices processed from different batches of single-cell RNA sequencing data obtained from different tissues, and one sample corresponds to the gene expression matrix processed from one batch of single-cell RNA sequencing data from one tissue; the target domain dataset includes n t samples, all samples in the target domain dataset are gene expression matrices processed from single-cell RNA sequencing data of different batches obtained from the same tissue; the source domain dataset is different from the target domain dataset, n s and n t are all positive integers greater than 1, and n s and n t Not equal. The single-cell gene expression matrix is ​​a two-dimensional matrix where each row represents a single cell and each column represents a gene. The elements in the matrix represent the expression levels of the corresponding genes in the corresponding cells, including gene expression matrices from multiple single-cell RNA sequencing data.

[0032] The source domain dataset is processed using a self-paced learning mechanism to determine a source domain training dataset for each iteration of a mini-batch gradient descent method;

[0033] Obtaining the cell category of each single-cell RNA sequencing data corresponding to the source domain training dataset at each iteration number of the mini-batch gradient descent method according to the source domain training dataset at each iteration number of the mini-batch gradient descent method and the cell category of each single-cell RNA sequencing data corresponding to the source domain dataset;

[0034] With the goal of minimizing the value of the total loss function, the scDRLN model is trained using the mini-batch gradient descent method according to the source domain training dataset at each iteration of the mini-batch gradient descent method, the target domain dataset, and the cell categories of each single-cell RNA sequencing data corresponding to the source domain training dataset at each iteration of the mini-batch gradient descent method to obtain a trained scDRLN model, wherein the first branch in the trained scDRLN model is used to determine the cell category; the total loss function includes a self-paced regularization term, a cross entropy loss function, a target domain dataset reconstruction loss function, a global alignment loss function, a source domain dataset alignment loss function, and a semantic centroid alignment loss function. ; The cross entropy loss function is determined based on the source domain training dataset, the target domain dataset and the first branch; the target domain dataset reconstruction loss function is determined based on the target domain dataset and the third branch; the global alignment loss function is determined based on the source domain training dataset, the target domain dataset and the second branch; the source domain dataset alignment loss function is determined based on the source domain training dataset and the feature encoder; the semantic centroid alignment loss function is determined based on the source domain training dataset, the target domain dataset, the label classifier and the feature encoder; the self-paced regularization term is determined according to the global alignment loss function.

[0035] In practical applications, the total loss function is: Among them, L total represents the total loss function, v i (λ,L d ) represents the sample selection variable, L l represents the cross entropy loss function, L r represents the target domain dataset reconstruction loss function, L d represents the global alignment loss function, L sou-c represents the source domain dataset alignment loss function, L sm represents the semantic centroid alignment loss function, f(v i ,λ) represents the self-step regularization term,γ represents the first balance parameter,δ represents the second balance parameter,α represents the third balance parameter,β represents the fourth balance parameter,n s It represents the number of cells corresponding to the single-cell RNA sequencing data corresponding to the source domain dataset, and λ represents the set threshold.

[0036] In practical applications, before obtaining the value of the global alignment loss function based on the source domain dataset, the target domain dataset, and the second branch, the method further includes:

[0037] like Figure 2As shown, the source domain dataset and the target domain dataset are sequentially subjected to quality control operations, normalization operations, and highly variable gene selection operations to obtain a processed source domain dataset and a processed target domain dataset.

[0038] In practical applications, the source domain dataset is processed using a self-paced learning mechanism to determine the source domain training dataset for each iteration of the mini-batch gradient descent method:

[0039] One or more gene expression matrices of single-cell RNA sequencing data are selected from the source domain dataset as the source domain dataset to be trained.

[0040] The target domain dataset and the source domain dataset to be trained are input into the second branch to obtain the domain to which the single-cell RNA sequencing data in the target domain dataset belongs and the domain to which each single-cell RNA sequencing data in the source domain dataset to be trained belongs.

[0041] The domain to which the single-cell RNA sequencing data in the target domain dataset belongs and the domain to which each single-cell RNA sequencing data in the source domain dataset to be trained belong are input into the global alignment loss function to obtain the value of the global alignment loss function corresponding to the source domain dataset to be trained.

[0042] If the value of the global alignment loss function corresponding to the source domain dataset to be trained is less than a set threshold, the source domain dataset to be trained is determined to be the source domain training dataset under the first iteration. The set threshold will change with the number of iterations.

[0043] If the value of the global alignment loss function corresponding to the source domain dataset to be trained is greater than or equal to the set threshold, return to the step of "selecting one or more gene expression matrices of single-cell RNA sequencing data from the source domain dataset as the source domain dataset to be trained".

[0044] In practical applications, the process of determining the value of the cross entropy loss function is:

[0045] The target domain dataset and the source domain training dataset are input into the first branch to obtain the output cell category L(G(X)) of the single-cell RNA sequencing data corresponding to the target domain dataset. t )) and the output cell categories of each single-cell RNA sequencing data corresponding to the source domain training dataset

[0046] The output cell category L(G(X) of the single-cell RNA sequencing data corresponding to the target domain dataset t )), the output cell category of each single-cell RNA sequencing data corresponding to the source domain training dataset And the cell categories of each single-cell RNA sequencing data corresponding to the source domain training dataset Input the cross entropy loss function to obtain the value L of the cross entropy loss function l .

[0047] In practical applications, the process of determining the value of the target domain dataset reconstruction loss function is as follows:

[0048] The target domain dataset is input into the third branch to obtain a reconstructed target domain dataset.

[0049] The reconstructed target domain dataset and the target domain dataset are input into the target domain dataset reconstruction loss function to obtain a value of the target domain dataset reconstruction loss function.

[0050] In practical applications, the process of determining the value of the global alignment loss function is as follows:

[0051] The target domain dataset and the source domain training dataset are input into the second branch to obtain the domain to which each single-cell RNA sequencing data in the target domain dataset belongs and the domain to which each single-cell RNA sequencing data in the source domain training dataset belongs.

[0052] The domain to which the single-cell RNA sequencing data in the target domain dataset belongs and the domain to which each single-cell RNA sequencing data in the source domain training dataset belongs are input into the global alignment loss function to obtain a value of the global alignment loss function.

[0053] In practical applications, the process of determining the value of the source domain dataset alignment loss function is as follows:

[0054] The source domain training data set is input into the feature extractor to obtain the features of the gene expression matrix of each single-cell RNA sequencing data in the source domain training data set.

[0055] The known cell category label information in the source domain training dataset is used as a division criterion to determine the positive sample set and the negative sample set, wherein the positive samples are cell samples from the same category, and the negative samples are cell samples from different categories. In the embedding space, the contrast loss is defined by maximizing the distance between similar cell samples and minimizing the distance between dissimilar cell samples. Specifically, during the training process, for each cell sample, some positive samples and negative samples are usually selected. For the positive sample pairs, the model should map them to close embedding spaces, while for the negative sample pairs, the model should map them to distant embedding spaces. The features of the gene expression matrices of the two single-cell RNA sequencing data (one for the positive sample set and one for the negative sample set) in the divided dataset are input into the source domain dataset alignment loss function to obtain the value of the source domain dataset alignment loss function.

[0056] In practical applications, the process of determining the value of the semantic centroid alignment loss function is as follows:

[0057] The gene expression matrix of each single-cell RNA sequencing data corresponding to the source domain training dataset is divided into multiple source domain clusters according to the cell category of each single-cell RNA sequencing data corresponding to the source domain training dataset; one source domain cluster corresponds to one cell category.

[0058] For any cell category, the average value of the features of the gene expression matrix of all single-cell RNA sequencing data in the source domain cluster corresponding to the cell category is determined to be the centroid of the cell category corresponding to the source domain training dataset.

[0059] The target domain dataset is input into the feature extractor to obtain features of the gene expression matrix of the single-cell RNA sequencing data in the target domain dataset.

[0060] The gene expression matrix of the single-cell RNA sequencing data corresponding to the target domain dataset is divided into multiple target domain clusters according to the cell category of the single-cell RNA sequencing data corresponding to the target domain dataset; one target domain cluster corresponds to one cell category.

[0061] For any cell category, the average value of the features of the gene expression matrix of all single-cell RNA sequencing data in the target domain cluster corresponding to the cell category is determined to be the centroid of the cell category corresponding to the target domain dataset.

[0062] The centroids of each cell category corresponding to the source domain training dataset and the centroids of each cell category corresponding to the target domain dataset are input into the semantic centroid alignment loss function to obtain the value of the semantic centroid alignment loss function.

[0063] In practical applications, the process of determining the value of the self-paced regularization term is:

[0064] According to the formula Calculate, where f(v i ,λ) represents the self-paced regularization term, λ represents the set threshold, n s Indicates the number of cells corresponding to the single-cell RNA sequencing data corresponding to the source domain dataset, v i (λ,L d ) represents the i-th sample selection variable.

[0065] The present invention provides a specific embodiment to explain in detail the idea of ​​the cell type classification method provided in the above embodiment:

[0066] The cell type classification method disclosed in the embodiment of the present invention includes a self-paced learning sample selection module (①), feature extraction (②), classification module (③), domain adaptive global alignment module (④) and local semantic alignment module (⑤). The overall framework structure of the cell type classification method is as follows Figure 1 As shown in Figure 1. It consists of a self-paced learning mechanism, a feature extractor G, a decoder G', a domain discriminator D, and a label classifier L. The self-paced learning mechanism allows the model to gradually learn from simple samples to complex samples, thereby better understanding the data distribution and characteristics, improving the robustness and generalization ability of the model, and thus improving the performance of cell category classification. The role of the feature extractor is to reduce the feature dimension in the dataset and further learn the effective representation of the features. The target domain dataset reconstruction loss function L r The purpose is to retain the feature information of the target domain data as much as possible, thereby improving the identifiability of the target domain features. The label classifier is used to predict the cell category of the input sample and assign pseudo labels to the samples from the target domain. The cross entropy loss function BatchNuclear-norm Maximization is used to improve the quality of the pseudo labels. The role of the domain discriminator is to distinguish whether the feature information comes from the source domain dataset or the target domain dataset. When the discriminator cannot correctly distinguish the source of the extracted features, it can learn the feature representation common to the source domain and the target domain, thereby achieving the purpose of globally confusing batch information. The source domain dataset alignment loss function L sou-c The role of is to make the intra-class distance of cells of the same category in the source domain closer, and the semantic centroid alignment loss function L sm The purpose of is to align the centroids of cell clusters of the same category in the source and target domains, thereby further correcting the batch effect at a fine-grained level and improving the accuracy of cell category classification.

[0067] This embodiment sets up a fully connected neural network layer, where the feature extractor G is composed of two hidden layers of 512 and 256 nodes, the hidden layer of the label classifier L is 256 nodes, and the softmax is used as the activation function in the last layer. The hidden layer of the domain discriminator D is set to 1024 nodes, using RELU as the activation function, and the sigmoid function is applied in the last layer. The hidden layer consists of a fully connected layer.

[0068] Step 1: Obtain single-cell RNA sequencing data of the source domain and target domain, preprocess the data, and obtain the preprocessed gene expression matrix.

[0069] The datasets used (source domain dataset and target domain dataset) are preprocessed with the help of Python's Scanpy package.

[0070] First, quality control of the dataset used was performed.

[0071] Use the scanpy.pp.filter_cells() function to filter out lower-quality cell samples in the dataset used, and use the scanpy.pp.filter_genes() function to filter out low-expressed genes in the dataset used. Then use string.startswith('MT-') to annotate mitochondrial genes (that is, gene names, because the names of mitochondrial-encoded genes start with the prefix mt- or MT-), and use the scanpy.pp.calculate_qc_metrics() function to calculate the number of mitochondrial genes in each cell, filtering out cells with highly expressed mitochondrial genes.

[0072] Then, the gene expression matrix was normalized to obtain the gene expression matrix X; finally, differentially expressed genes were selected for downstream analysis.

[0073] Specifically, the count of each cell is standardized using the scanpy.pp.normalize_total() function, and then logarithmically transformed using scanpy.pp.log1p().

[0074] Finally, scanpy.pp.highly_variable_genes() function is used to select highly variable genes and save them as the preprocessed gene expression matrix X, which can be used for downstream analysis, such as cell class annotation.

[0075] Step 2: Based on the gene expression matrix X in step 1, define the source domain dataset and the target domain dataset, and use the reconstruction loss to constrain the extraction of target domain features to obtain the features embedded in the source domain and target domain in the low-dimensional space.

[0076] Let the source domain dataset be The corresponding n s The label information is recorded as Indicates the nth s The gene expression matrix of single-cell RNA sequencing data, X s Each single-cell RNA sequencing data (one sample) has p highly variable gene features. Represents the cell category of the i-th single-cell RNA sequencing data in the source domain dataset, and the target domain dataset is recorded as No label information, Indicates the nth t Gene expression matrix of single-cell RNA sequencing data.

[0077] The batch information b in the dataset is associated with a specific domain D, and b and D are used interchangeably in the text (here, the batch is actually equivalent to the domain, for example, a batch is a domain, such as the batch information b of the source domain dataset s and source domain D s , is just an equivalence relationship, referring to b s It is equivalent to the source domain D s ). In fact, in a dataset with batch effects, the distribution of cell classes is related to their batch information, so X s and X t The distributions in the source domain D are different but related. s Train a specific batch of samples in the target domain D t Other specific batches of samples were tested, where b s ≠b t , b t is another specific batch in the dataset that is different from the batch in the source domain dataset.

[0078] In order to better preserve the local features of the target domain dataset, reduce the loss of target domain feature information, and align the source and target domain data distributions, the extracted target domain features are made more discriminative. Therefore, the feature information obtained by the feature extractor is decoded and reconstructed, and MSE Loss is used as the reconstruction loss function of the target domain dataset, that is, the target domain dataset reconstruction loss function L r , the specific formula is as follows:

[0079]

[0080] in, The target domain dataset is input into the feature extractor and then decoded into the decoder, n t The total number of single-cell RNA sequencing data corresponding to the target domain dataset is minimized by mapping the input to a feature space (this is the feature extraction process), and then mapping it back from this feature space to the input space for reconstruction (this is reconstruction). This can better preserve the characteristics of the target domain dataset.

[0081] Step 3: Train a label classifier using the source domain dataset from step 2.

[0082] Each row of data in the source domain dataset Represents a sample. Here, the features of the extracted source domain dataset are put into the label classifier to classify the output cell categories of each single-cell RNA sequencing data corresponding to the source domain dataset.

[0083] Use Cross-entropy Loss (cross loss function) as the classification loss of the source domain dataset, that is, L l The first item, and the classifier is used in the local distribution alignment stage to assign pseudo labels to the target domain dataset, that is, output cell categories. In order to ensure the quality of the generated pseudo labels, in order to obtain reliable pseudo labels, Batch Nuclear-norm Maximization (BNM) is introduced as the cross entropy loss function L after the constraint / penalty term. l , the specific formula is as follows:

[0084]

[0085] in, is the category reweighting coefficient, which is based on the number of different cell categories in the source domain dataset and gives different weights to each category to balance the difference in the amount of data of different categories. for The corresponding output cell category of single-cell RNA sequencing data, for Features, coming soon The input feature extractor is also called the embedding representation of the source domain dataset in the low-dimensional space, K is the number of cell categories, L(G(X t )) is X t The corresponding output cell category of the single-cell RNA sequencing data is the predicted output matrix of the target domain data, G(X t ) is X t In fact, the cross entropy classification loss L l What is calculated is the negative logarithm of the classification confidence corresponding to the actual cell category to which each cell belongs. In other words, it is how likely the model prediction is correct. l The second term is the BNM nuclear norm, which is used to enhance the diversity and discriminability of predictions after domain adaptation.

[0086] Step 4: Use the domain discriminator to confuse samples from the source and target domains.

[0087] Global alignment is achieved based on domain discriminator. Feature information from source and target domains is trained adversarially, and binary cross entropy loss, i.e. global alignment loss function L, is used. d Evaluate the game between the source domain and the target domain. The specific formula is as follows:

[0088]

[0089] in, for The features of the target domain dataset are obtained by inputting the feature extractor, which is also called the embedding representation of the target domain dataset in the low-dimensional space. A gradient reversal layer is set between G and D to flip the gradient sign between G and D during back propagation in order to maximize the domain loss and thus achieve the purpose of confusing the discriminator. In short, the task of the domain discriminator is to distinguish whether two samples are from different batches, so that the feature representation generated by the G network cannot be used for classification based on batch information, thereby weakening the batch effect between the source domain and the target domain, making L d The domain discriminator is trained to the minimum value, so that the domain discriminator is not correctly classified, which means that the global alignment is achieved, and the domain deviation is reduced, thereby achieving the purpose of weakening the batch effect.

[0090] Step 5: Compact the source domain dataset based on contrastive loss.

[0091] To further correct for batch effects, this step and step 7 are both local corrections. The true labels of the source domain are used. The input is the features of the gene expression matrix extracted from the source domain dataset and the labels of the source domain dataset. The result is that the distances between similar samples are more compact.

[0092] Align the source domain dataset based on existing labels. For single-cell RNA sequence data, when performing cell category classification, it is usually a multi-classification task. Under the premise of global domain invariance, it is hoped that the sample set can be compact within the class and separable between classes in the feature space. Since the source domain dataset carries label information, the Contrastive Loss, that is, the source domain dataset alignment loss function L, is used. sou-c The distance between cells of the same category in the source domain is shortened, and the distance between cells of different categories is pushed away. The Contrastive Loss formula is as follows:

[0093]

[0094] in, and is the feature of the gene expression matrix of two single-cell RNA sequencing data randomly selected in the source domain dataset, and E(·) represents and The Euclidean distance, U = 1 means that the two samples belong to the same cell category, while U = 0 means they belong to different cell categories. m is the threshold of the prediction boundary, and L sou-c At the very least, the distance between cells of the same category can be shortened, while the distance between cells of different categories can be pushed further away.

[0095] Step 6: Align the centroids of the same category cell clusters in the source and target domains using semantic centroid loss based on pseudo labels.

[0096] Based on the trained label classifier, pseudo labels are assigned to the target domain dataset, and semantic centroid alignment loss is used to locally align samples of the same category in the source and target domains. In order to ensure the distinguishability of embeddings of cells of the same category in different batches, it is decided to explicitly align cell clusters of the same category in the source and target domains at the class level. Since the target domain lacks label information of cell categories, the label classifier L trained in the source domain is used to assign pseudo labels to the target domain dataset. For incorrect labels in the target domain, semantic centroid alignment is used to weaken the deviation caused by some incorrect labels. When all pseudo labels are used for centroid calculation, the correct pseudo labels are expected to offset the impact of incorrect labels. Therefore, Semantic Center alignment Loss, that is, semantic centroid alignment loss function L, is used. sm Minimize the centroids of the same type of cell clusters in the source and target domains, and align them. This not only suppresses the negative impact of incorrect pseudo-labels, but also makes different types of cell clusters separable. The Semantic Center Alignment Loss formula is as follows:

[0097]

[0098] in, It represents the centroid of the k-th cell category in the source domain dataset. It represents the centroid of the k-th cell category in the target domain dataset.

[0099] Step 7: Finalize the cell categories of the target domain dataset and complete the classification. This step is to jointly optimize the network and determine the final pseudo-label results, that is, the final cell category classification results.

[0100] The loss L of the scDRLN model is defined as follows:

[0101] L=L l +γL r +δL d +αL sou-c +βL sm

[0102] Here μ is a hyperparameter, δ, γ, α, and β are balance parameters, which represent the contribution of global domain alignment loss, target domain dataset reconstruction loss, source domain dataset alignment loss, and semantic centroid alignment loss to the total loss, respectively. Set δ = μ, γ = 0.1μ, α = 0.1μ, β = μ, where the parameter μ is specified according to the following formula:

[0103]

[0104] Here, t is a parameter that increases linearly from 0 to 1 and has no specific physical meaning.

[0105] The parameters of the scDRLN model are optimized by minimizing the loss function L. During training, gradients are calculated and the parameters of the scDRLN model are updated using the backpropagation algorithm. As training progresses, the scDRLN model will gradually adjust its parameters to reduce the loss function, thereby improving the accuracy of cell type label prediction.

[0106] Model evaluation: Using the target domain dataset X t Evaluate the trained model. Input the test sample into the model to obtain the model's prediction. Typically, the prediction result is a probability distribution vector, representing the probability of each class. To determine the final predicted label, select the class with the highest probability as the prediction result. For example, if the probability distribution vector of the prediction result is [0.1, 0.2, 0.6, 0.05, 0.05], select the third class with the highest probability as the predicted label.

[0107] Step 8: Select the source domain dataset for the input model based on the self-paced learning mechanism.

[0108] In order to alleviate the imbalance of sample categories and suppress the negative impact of pseudo labels, a self-paced learning mechanism is introduced to control the model to first learn samples that are similar between domains, and then increase the difficulty to gradually learn samples with large differences between domains. Ultimately, all source domain datasets will be trained. Here, the parameter v is used. i (λ,L d ) control learning, v i (λ,L d ) can take 0 and 1, indicating the difficulty of the sample. When vi(λ,L d )=1, it means that the sample is selected for training in the next iteration, and v i (λ,L d )=0, it means that the sample is not selected for training. Calculate v i (λ,L d ), λ is a hyperparameter, v i (λ,L d ) by f(v i ,λ) determines whether to select this sample, where f(v i ,λ) is v i (λ,L d ) is a decreasing function of λ,

[0109] The total loss L of scDRLN after the introduction of the self-paced learning mechanism total The definition is as follows:

[0110]

[0111] The scDRLN network is trained using mini-batch gradient descent with momentum set to 0.9 and weight decay to 5×10 -4 , following η p =η0 / (1+ap) q The learning rate is dynamically adjusted in this way, where p increases linearly from 0 to 1, η0 is the initial learning rate, set to 0.001, a=10, q=0.75, and the batch size is 256.

[0112] In summary, the scDRLN model provided by the embodiment of the present invention is a domain adaptive network based on representation learning, which can simultaneously correct batches and accurately classify cell categories. Figure 3 As shown, the present invention first uses the Scanpy package to preprocess the gene expression matrix (corresponding to S1); then uses the encoder to extract features from the preprocessed samples (corresponding to S2), and trains a class-reweighted classifier in the source domain to classify the cell categories (corresponding to S3). Secondly, a domain classifier is trained between domains to confuse the source domain dataset and the target domain dataset, so that the encoder learns the invariant feature representation of the source domain and the target domain, corrects the batch effect to a certain extent, and improves the accuracy of cell category classification (corresponding to S4); then, the samples in the source domain are centrally aligned using contrast loss to eliminate the differences between them (corresponding to S5), and samples from different domains but with the same class label are locally aligned based on the pseudo-labels assigned by the source domain classifier, further eliminating the batch effect and improving the cell category classification performance (corresponding to S6), and finally the cell category of the target domain dataset is determined, and correction and classification are completed (corresponding to S7). Finally, based on the inter-domain loss, a self-stepping learning method is used to gradually select a source domain dataset similar to the target domain for training, thereby improving the performance of the model (corresponding to S8).

[0113] An embodiment of the present invention provides a cell type determination system corresponding to the above method, the system comprising:

[0114] A model construction module is used to construct a scDRLN model; the scDRLN model includes a first branch, a second branch and a third branch; the first branch includes a feature extractor and a label classifier connected in sequence; the second branch includes a feature extractor and a domain discriminator connected in sequence; the third branch includes a feature extractor and a decoder connected in sequence.

[0115] The acquisition module is used to obtain the source domain dataset, the target domain dataset and the cell categories of each single cell RNA sequencing data corresponding to the source domain dataset; the source domain dataset includes n ssamples; all samples in the source domain dataset are gene expression matrices processed from different batches of single-cell RNA sequencing data obtained from different tissues; the target domain dataset includes n t samples, all samples in the target domain dataset are gene expression matrices processed from single-cell RNA sequencing data of different batches obtained from the same tissue; the source domain dataset is different from the target domain dataset, n s and n t are all positive integers greater than 1, and n s and n t Not equal.

[0116] The self-paced learning mechanism module is used to process the source domain data set using the self-paced learning mechanism to determine the source domain training data set under each iteration number of the small batch gradient descent method.

[0117] The training data cell category determination module is used to obtain the cell category of each single-cell RNA sequencing data corresponding to the source domain training dataset under each iteration number of the mini-batch gradient descent method according to the source domain training dataset under each iteration number of the mini-batch gradient descent method and the cell category of each single-cell RNA sequencing data corresponding to the source domain dataset.

[0118] The training module is used to train the scDRLN model with the goal of minimizing the value of the total loss function, according to the source domain training dataset under each iteration of the mini-batch gradient descent method, the target domain dataset, and the cell category of each single-cell RNA sequencing data corresponding to the source domain training dataset under each iteration of the mini-batch gradient descent method, to obtain a trained scDRLN model, wherein the first branch in the trained scDRLN model is used to determine the cell category; the total loss function includes a self-paced regularization term, a cross entropy loss function, a target domain dataset reconstruction loss function, a global alignment loss function, a source domain dataset alignment loss function, and a semantic centroid alignment loss function. loss function; the cross entropy loss function is determined based on the source domain training dataset, the target domain dataset and the first branch; the target domain dataset reconstruction loss function is determined based on the target domain dataset and the third branch; the global alignment loss function is determined based on the source domain training dataset, the target domain dataset and the second branch; the source domain dataset alignment loss function is determined based on the source domain training dataset and the feature encoder; the semantic centroid alignment loss function is determined based on the source domain training dataset, the target domain dataset, the label classifier and the feature encoder; the self-paced regularization term is determined according to the global alignment loss function.

[0119] The present invention has the following beneficial effects:

[0120] 1) This paper employs domain adaptation and representation learning-based techniques to integrate the correction and classification processes, achieving mutually reinforcing and complementary results. This jointly optimized two-stage alignment is achieved by aligning global distributions and category-level features. This not only effectively learns the semantic information of the samples, but also simultaneously improves batch effect correction and cell category classification accuracy.

[0121] 2) The present invention can effectively retain the local features of the target domain dataset through the autoencoder, and can improve the accuracy of pseudo labels in the target domain by using BatchNuclear-norm Maximization. At the same time, the introduction of self-paced learning and category-level reweighting mechanism helps to eliminate the negative impact of difficult-to-classify samples on the initial model and alleviate the problem of sample category imbalance in the dataset, further enhancing the robustness and generalization ability of the model.

[0122] 3) The present invention adopts contrast loss to perform feature alignment on the source domain dataset, and uses semantic alignment loss to align the centroids of cell clusters of the same category in the source and target domains based on pseudo-labels. This helps to make cell clusters of the same category more compact, cell clusters of different categories easier to distinguish, and helps to further correct batch effects.

[0123] 4) Experimental results on cross-library and cross-platform datasets show that this method outperforms current advanced methods in batch effect correction and cell category classification, and the effectiveness of this method is verified through ablation studies.

[0124] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0125] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A method for determining cell type, characterized in that: include: Constructing a scDRLN model; the scDRLN model includes a first branch, a second branch, and a third branch; The first branch includes a feature extractor and a label classifier connected in sequence; the second branch includes a feature extractor and a domain discriminator connected in sequence; the third branch includes a feature extractor and a decoder connected in sequence; Obtain a source domain dataset, a target domain dataset, and the cell categories of each single-cell RNA sequencing data corresponding to the source domain dataset; the source domain dataset includes n s samples; all samples in the source domain dataset are gene expression matrices processed from different batches of single-cell RNA sequencing data obtained from different tissues; the target domain dataset includes n t samples, all samples in the target domain dataset are gene expression matrices processed from single-cell RNA sequencing data of different batches obtained from the same tissue; the source domain dataset is different from the target domain dataset, n s and n t are all positive integers greater than 1, and n s and n t Not equal; The source domain dataset is processed using a self-paced learning mechanism to determine a source domain training dataset for each iteration of a mini-batch gradient descent method; Obtaining the cell category of each single-cell RNA sequencing data corresponding to the source domain training dataset at each iteration number of the mini-batch gradient descent method according to the source domain training dataset at each iteration number of the mini-batch gradient descent method and the cell category of each single-cell RNA sequencing data corresponding to the source domain dataset; With the goal of minimizing the value of the total loss function, the scDRLN model is trained using the mini-batch gradient descent method according to the source domain training dataset at each iteration of the mini-batch gradient descent method, the target domain dataset, and the cell categories of each single-cell RNA sequencing data corresponding to the source domain training dataset at each iteration of the mini-batch gradient descent method to obtain a trained scDRLN model, wherein the first branch in the trained scDRLN model is used to determine the cell category; the total loss function includes a self-paced regularization term, a cross entropy loss function, a target domain dataset reconstruction loss function, a global alignment loss function, a source domain dataset alignment loss function, and a semantic centroid alignment loss function; the cross entropy loss function is determined based on the source domain training dataset, the target domain dataset, and the first branch; the target domain dataset reconstruction loss function is determined based on the target domain dataset and the third branch; The global alignment loss function is determined based on the source domain training dataset, the target domain dataset and the second branch; The source domain dataset alignment loss function is determined based on the source domain training dataset and the feature encoder; The semantic centroid alignment loss function is determined based on the source domain training dataset, the target domain dataset, the label classifier and the feature encoder; the self-paced regularization term is determined according to the global alignment loss function.

2. The method for determining cell types according to claim 1, wherein: The total loss function is: Among them, L total Represents the total loss function, v i (λ,L d ) represents the i-th sample selection variable, L l represents the cross entropy loss function, L r represents the target domain dataset reconstruction loss function, L d represents the global alignment loss function, L sou-c represents the source domain dataset alignment loss function, L sm represents the semantic centroid alignment loss function, f(v i ,λ) represents the self-step regularization term,γ represents the first balance parameter,δ represents the second balance parameter,α represents the third balance parameter,β represents the fourth balance parameter,n s It represents the number of cells corresponding to the single-cell RNA sequencing data corresponding to the source domain dataset, and λ represents the set threshold.

3. The method for determining cell type according to claim 1, wherein: Before obtaining the value of the global alignment loss function according to the source domain dataset, the target domain dataset, and the second branch, the method further includes: The source domain dataset and the target domain dataset are sequentially subjected to quality control operations, normalization operations, and highly variable gene selection operations to obtain a processed source domain dataset and a processed target domain dataset.

4. The method for determining cell type according to claim 1, wherein: The process of determining the value of the cross entropy loss function is: Inputting the target domain dataset and the source domain training dataset into the first branch to obtain the output cell category of the single-cell RNA sequencing data corresponding to the target domain dataset and the output cell category of each single-cell RNA sequencing data corresponding to the source domain training dataset; The output cell category of the single-cell RNA sequencing data corresponding to the target domain dataset, the output cell category of each single-cell RNA sequencing data corresponding to the source domain training dataset, and the cell category of each single-cell RNA sequencing data corresponding to the source domain training dataset are input into the cross entropy loss function to obtain the value of the cross entropy loss function.

5. The method for determining cell types according to claim 4, wherein: The process of determining the value of the target domain dataset reconstruction loss function is as follows: Inputting the target domain dataset into the third branch to obtain a reconstructed target domain dataset; The reconstructed target domain dataset and the target domain dataset are input into the target domain dataset reconstruction loss function to obtain a value of the target domain dataset reconstruction loss function.

6. The method for determining cell types according to claim 5, wherein: The process of determining the value of the global alignment loss function is as follows: Inputting the target domain dataset and the source domain training dataset into the second branch to obtain the domain to which the single-cell RNA sequencing data in the target domain dataset belongs and the domain to which each single-cell RNA sequencing data in the source domain training dataset belongs; The domain to which the single-cell RNA sequencing data in the target domain dataset belongs and the domain to which each single-cell RNA sequencing data in the source domain training dataset belongs are input into the global alignment loss function to obtain a value of the global alignment loss function.

7. The method for determining cell types according to claim 6, wherein: The process of determining the value of the source domain dataset alignment loss function is as follows: Inputting the source domain training data set into the feature extractor to obtain features of the gene expression matrix of each single-cell RNA sequencing data in the source domain training data set; Inputting the features of the gene expression matrices of any two single-cell RNA sequencing data in the source domain training dataset into the source domain dataset alignment loss function to obtain a value of the source domain dataset alignment loss function; One of the gene expression matrices of any two single-cell RNA sequencing data comes from a positive sample set, and the other comes from a negative sample set; the positive sample set and the negative sample set are obtained according to the cell categories of each single-cell RNA sequencing data corresponding to the source domain training dataset.

8. The method for determining cell types according to claim 7, wherein: The process of determining the value of the semantic centroid alignment loss function is as follows: Dividing the gene expression matrix of each single-cell RNA sequencing data corresponding to the source domain training dataset into multiple source domain clusters according to the cell category of each single-cell RNA sequencing data corresponding to the source domain training dataset; one source domain cluster corresponds to one cell category; For any cell category, determining that the average value of the features of the gene expression matrix of all single-cell RNA sequencing data in the source domain cluster corresponding to the cell category is the centroid of the cell category corresponding to the source domain training dataset; Inputting the target domain dataset into the feature extractor to obtain features of the gene expression matrix of the single-cell RNA sequencing data in the target domain dataset; Dividing the gene expression matrix of the single-cell RNA sequencing data corresponding to the target domain dataset into multiple target domain clusters according to the cell category of the single-cell RNA sequencing data corresponding to the target domain dataset; one target domain cluster corresponds to one cell category; For any cell category, determining that the average value of the features of the gene expression matrix of all single-cell RNA sequencing data in the target domain cluster corresponding to the cell category is the centroid of the cell category corresponding to the target domain dataset; The centroids of each cell category corresponding to the source domain training dataset and the centroids of each cell category corresponding to the target domain dataset are input into the semantic centroid alignment loss function to obtain the value of the semantic centroid alignment loss function.

9. The method for determining cell types according to claim 8, wherein: The process of determining the value of the self-paced regularization term is: According to the formula Calculate, where f(v i ,λ) represents the self-paced regularization term, λ represents the set threshold, n s Indicates the number of cells corresponding to the single-cell RNA sequencing data corresponding to the source domain dataset, v i (λ,L d ) represents the i-th sample selection variable.

10. A cell type determination system, characterized in that: include: A model construction module for constructing a scDRLN model; the scDRLN model includes a first branch, a second branch, and a third branch; the first branch includes a feature extractor and a label classifier connected in sequence; the second branch includes a feature extractor and a domain discriminator connected in sequence; the third branch includes a feature extractor and a decoder connected in sequence; The acquisition module is used to obtain the source domain dataset, the target domain dataset and the cell categories of each single cell RNA sequencing data corresponding to the source domain dataset; the source domain dataset includes n s samples; all samples in the source domain dataset are gene expression matrices processed from different batches of single-cell RNA sequencing data obtained from different tissues; the target domain dataset includes n t samples, all samples in the target domain dataset are gene expression matrices processed from single-cell RNA sequencing data of different batches obtained from the same tissue; the source domain dataset is different from the target domain dataset, n s and n t are all positive integers greater than 1, and n s and n t Not equal; A self-paced learning mechanism module, configured to process the source domain dataset using a self-paced learning mechanism to determine a source domain training dataset for each iteration of a small batch gradient descent method; a training data cell category determination module, configured to obtain the cell category of each single-cell RNA sequencing data corresponding to the source domain training dataset at each iteration number of the mini-batch gradient descent method according to the source domain training dataset at each iteration number of the mini-batch gradient descent method and the cell category of each single-cell RNA sequencing data corresponding to the source domain dataset; A training module, for training the scDRLN model using a mini-batch gradient descent method with the goal of minimizing the value of the total loss function, according to the source domain training dataset at each iteration of the mini-batch gradient descent method, the target domain dataset, and the cell category of each single-cell RNA sequencing data corresponding to the source domain training dataset at each iteration of the mini-batch gradient descent method, to obtain a trained scDRLN model, wherein the first branch in the trained scDRLN model is used to determine the cell category; the total loss function includes a self-paced regularization term, a cross-entropy loss function, a target domain dataset reconstruction loss function, a global alignment loss function, a source domain dataset alignment loss function, and a semantic centroid alignment loss function; the cross-entropy loss function is determined based on the source domain training dataset, the target domain dataset, and the first branch; the target domain dataset reconstruction loss function is determined based on the target domain dataset and the third branch; The global alignment loss function is determined based on the source domain training dataset, the target domain dataset and the second branch; The source domain dataset alignment loss function is determined based on the source domain training dataset and the feature encoder; The semantic centroid alignment loss function is determined based on the source domain training dataset, the target domain dataset, the label classifier and the feature encoder; the self-paced regularization term is determined according to the global alignment loss function.

Images