A single-cell data deep clustering method and device based on a double autoencoder and a medium

Abstract

The application discloses a single-cell data deep clustering method and device based on a double self-encoder and a medium, and relates to the technical field of single-cell RNA sequencing. The method comprises the following steps: acquiring original gene data of single cells; constructing an scDAEC model; the scDAEC model comprises a dynamic combination multi-head attention mechanism, a denoising self-encoder and a graph self-encoder; the denoising self-encoder is constructed based on a zero-inflated negative binomial distribution loss adjusted by a zero inflation proportion adaptive weight; the graph self-encoder is constructed based on a graph convolution network; and the original gene data is input into the scDAEC model for clustering processing, so that a single-cell data clustering result is obtained. The framework constructed by the application has strong generalization ability and high fault tolerance.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of single-cell RNA sequencing, in particular to a single-cell data deep clustering method based on a double autoencoder, equipment and medium. BACKGROUND

[0002] Single-cell sequencing is an important tool for studying cell heterogeneity and developmental processes, and has greatly promoted the progress in the fields of biology and medicine. Single-cell RNA sequencing (scRNA-seq) technology has high resolution and sensitivity by analyzing gene expression of single cells, and can accurately depict different cell types and states in tissues. Single-cell clustering is an important step in analyzing single-cell data, which can divide cells with the same or similar characteristics into several categories according to the similarity between cells, and identify different cells in the organism, providing key support for analyzing the complexity of biological systems.

[0003] It can be seen that single-cell RNA sequencing technology has become an important means for studying gene expression, cell heterogeneity and complex biological processes at the single-cell level. Clustering analysis, as a crucial step in scRNA-seq data processing, is of great significance for identifying subtype differences of cell types. Due to the sparsity, high dimensionality and noise of scRNA-seq data, traditional clustering methods are difficult to effectively process. In recent years, deep learning has shown potential in single-cell clustering, but existing methods focus on local feature learning and ignore gene correlation, limiting clustering effectiveness. At the same time, a large number of zero values make the model easily misjudge the true low expression genes as technical zeros, affecting key biological information. Therefore, how to fully utilize the dependency between genes and integrate the contributions of different expression level genes is a key challenge to improve clustering accuracy. SUMMARY

[0004] The purpose of the present application is to provide a single-cell data deep clustering method based on a double autoencoder, equipment and medium, which aims to solve or improve at least one of the above technical problems.

[0005] To achieve the above purpose, the present application provides the following solutions:

[0006] A single-cell data deep clustering method based on a double autoencoder, comprising:

[0007] obtaining original gene data of single cells;

[0008] Construct an scDAEC model; the scDAEC model includes a dynamically combined multi-head attention mechanism, a denoising autoencoder, and a graph autoencoder; the denoising autoencoder is constructed based on a zero-inflation negative binomial distribution loss adjusted by zero-inflation proportional adaptive weights; the graph autoencoder is constructed based on a graph convolutional network;

[0009] The original gene data is input into the scDAEC model for clustering to obtain single-cell data clustering results.

[0010] Optionally, the step of inputting the original gene data into the scDAEC model for clustering processing to obtain single-cell data clustering results specifically includes:

[0011] The raw gene data is preprocessed and sorted according to gene content to generate a gene expression matrix; the rows in the gene expression matrix represent cells and the columns represent genes.

[0012] A directed graph is constructed based on the KNN algorithm and the gene expression matrix;

[0013] The directed graph is input into the denoising autoencoder and the graph autoencoder respectively. In the denoising autoencoder, topological embeddings are generated by balancing dropout and real low-expression genes. In the graph autoencoder, the topological relationships between cells are captured to generate denoising embeddings.

[0014] By combining the topological embedding and the denoising embedding using a dynamic multi-head attention mechanism, single-cell data clustering results are obtained.

[0015] Optionally, the preprocessing includes quality control, standardization, logarithmic transformation, and normalization.

[0016] Optionally, the specific processing procedure in the graph autoencoder is as follows:

[0017] A local aggregation mechanism is used to smooth noise in the data, and the following computations are performed in each layer of the graph convolutional network:

[0018] ,

[0019] in, It represents a symmetric normalized adjacency matrix with self-loops added to ensure that information can propagate in the network; It is the first Layer node feature embedding represents a low-dimensional representation of a cell; It is the first The trainable weight matrix of the layer, where tanh represents the hyperbolic tangent function;

[0020] In the decoding part, the dot product reconstruction method is used to calculate the dot product reconstructed adjacency matrix of the node embedding:

[0021] ,

[0022] in, It is the reconstructed adjacency matrix. It uses the Sigmoid activation function; finally, the mean squared error loss function is established to measure the error between the reconstructed image and the original image:

[0023] ,

[0024] in, and The original graph and the reconstructed graph respectively represent the first two graphs in the first two graphs. row and number The elements of the column are optimized by minimizing the mean squared error loss function to obtain a topological embedding that conforms to the intercellular relationships.

[0025] Optionally, the step of combining the topological embedding and the denoised embedding using a dynamic multi-head attention mechanism to obtain single-cell data clustering results specifically includes:

[0026] First, the topological embedding and the denoised embedding are projected onto the query, key, and value spaces, respectively. Second, the attention score is calculated using the dot product attention mechanism and normalized using the softmax function to convert the attention score into a probability distribution. Then, the weight of each attention head is dynamically calculated based on the input data. Finally, the outputs of each attention head are summed using dynamic weights to obtain the single-cell data clustering result.

[0027] Optionally, the zero-inflated negative binomial distribution loss adjusted by the zero-inflated proportional adaptive weights is:

[0028] ,

[0029] in, For adaptive weights, The weights for the zero-inflation portion, Indicates an indicator function, N This represents the total number of elements in the dataset. This represents the probability of the zero-inflation portion. This represents the negative binomial distribution probability mass function. Represents cells In genes The original count on, The parameter representing the mean of the negative binomial distribution. Represents discrete parameters.

[0030] The present invention also provides an electronic device, including a memory and a processor, wherein the memory is used to store a computer program, and the processor runs the computer program to enable the electronic device to perform the single-cell data deep clustering method based on dual autoencoders described above.

[0031] The present invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the single-cell data deep clustering method based on a dual autoencoder as described above.

[0032] According to specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0033] This invention discloses a method, device, and medium for deep clustering of single-cell data based on a dual autoencoder. The method includes acquiring raw gene data of single cells; constructing a scDAEC model; the scDAEC model includes a dynamically combined multi-head attention mechanism, a denoising autoencoder, and a graph autoencoder; the denoising autoencoder is constructed based on a zero-inflated negative binomial distribution loss adjusted by zero-inflated proportional adaptive weights; the graph autoencoder is constructed based on a graph convolutional network; the raw gene data is input into the scDAEC model for clustering processing to obtain the single-cell data clustering results. The framework constructed in this invention has strong generalization ability and high fault tolerance. Attached Figure Description

[0034] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0035] Figure 1 This is the overall architecture diagram of scDAEC in this embodiment;

[0036] Figure 2 This is a comparison chart of ARI scores of scDAEC and ten other clustering methods on 13 datasets in this embodiment;

[0037] Figure 3 This is a comparison chart of the NMI scores of scDAEC and ten other clustering methods on 13 datasets in this embodiment;

[0038] Figure 4 This example uses the Muraro scDAEC dataset to visualize and compare it with four other deep learning models using UMAP.

[0039] Figure 5This is a UMAP visualization comparison chart of scDAEC using the Quake_10x_Limb_Muscle dataset and four other deep learning models in this embodiment;

[0040] Figure 6 This is the silhouette coefficient detection map obtained using scDAEC and four other clustering methods in this embodiment;

[0041] Figure 7 This is a Sankey diagram comparing the clustering results of scDAEC and several other methods on the Zilionis_mouse_lung dataset in this embodiment;

[0042] Figure 8 This is a comparison chart of the batch effect processing results of scDAEC and 8 other methods in this embodiment on 4 real datasets and 2 simulated datasets;

[0043] Figure 9 This example shows the cell trajectory plot and pseudo-time plot of the Adam dataset.

[0044] Figure 10 This is a comparison chart of the clustering results in this embodiment;

[0045] Figure 11 The clustering performance of scDAEC in this embodiment on eight scRNA-seq datasets (each dataset was tested five times) was given by dropout rates of 10%, 20%, 30%, 40%, and 50%.

[0046] Figure 12 This is a schematic diagram of the ARI scores of the scDAEC model after ablation experiments on 8 scRNA-seq datasets in this embodiment;

[0047] Figure 13 This is a schematic diagram of the NMI scores of the scDAEC model after ablation experiments on eight scRNA-seq datasets in this embodiment. Detailed Implementation

[0048] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0049] The purpose of this invention is to provide a method, device and medium for deep clustering of single-cell data based on dual autoencoders, aiming to solve or improve at least one of the above-mentioned technical problems.

[0050] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0051] like Figure 1 As shown, this invention provides a deep clustering method for single-cell data based on a dual autoencoder, comprising:

[0052] Step 100: Obtain the raw genetic data of a single cell.

[0053] Step 200: Construct the scDAEC model; the scDAEC model includes a dynamically combined multi-head attention mechanism, a denoising autoencoder, and a graph autoencoder; the denoising autoencoder is constructed based on a zero-inflation negative binomial distribution loss adjusted by zero-inflation proportional adaptive weights; the graph autoencoder is constructed based on a graph convolutional network.

[0054] Step 300: Input the original gene data into the scDAEC model for clustering processing to obtain single-cell data clustering results.

[0055] As a specific implementation method, the following embodiment is provided.

[0056] First, the raw data was acquired and preprocessed. This example used 13 real-world datasets summarized in Table 1, which can be downloaded from the NCBI Gene Expression Omnibus (GEO) database. To minimize data noise, the Scanpy package was used to preprocess the raw scRNA-seq data. scRNA-seq data is represented in matrix form, with rows representing cells and columns representing genes. First, quality control was performed to remove uncounted cells and genes, followed by normalization, logarithmic transformation, and standardization to improve data quality. In selecting highly variable genes, an appropriate number of genes can significantly reduce computational costs, avoid significant noise, and maintain high accuracy; most studies typically select 2000-3000 genes. During the experiment, it was found that the adjusted RAND Index (ARI) value was higher when the number of genes was around 2500; therefore, the number of highly variable genes was set to 2500. The genes were sorted according to their gene content for easier subsequent processing, forming the original counting matrix. Then, based on the processed matrix, a cell-cell directed graph is constructed using the KNN algorithm as input for GAE. Most methods select 10-20 nearest neighbors to construct the cell graph. In this embodiment, the similarity of the 15 nearest neighbors of each cell is calculated, and the similarity between cells is optimized using the Gaussian kernel function and cosine distance.

[0057] Table 1 Real Dataset

[0058] ,

[0059] Secondly, construct the overall architecture of scDAEC (such as...) Figure 1 As shown, the original gene expression matrix is ​​first subjected to quality control, standardization, logarithmic transformation, and normalization to obtain a matrix. Then, a directed graph is constructed based on the KNN algorithm, which is used as the input to GAE and DAE, respectively. The optimized DAE denoising autoencoder can balance dropout and real low-expression genes, avoiding the loss of biological information due to ignoring some real low-expression gene signals, and obtain denoised embeddings. At the same time, GAE captures the topological relationships between cells and learns the low-dimensional topological embeddings of graph structure nodes. Then, a dynamic combination multi-head attention mechanism is used to combine the denoised embeddings and topological embeddings. Finally, an unsupervised clustering algorithm is used for deep clustering and other downstream task analysis.

[0060] DAE based on adaptive ZINB loss:

[0061] For a given single-cell RNA dataset, its gene expression matrix is ​​represented as follows: ,in Represents cells, This represents the gene dimension of each cell. The encoder mainly consists of an encoder, a decoder, and a hidden layer. The encoder receives a series of preprocessed data. As input, the output data is obtained in the encoding layer. The decoder will then Map back to the original input through the hidden layer. The reconstructed data is obtained.

[0062] Both the encoder and decoder are based on multilayer perceptron (MLP) neural networks. MLPs are suitable for high-dimensional features of scRNA-seq data. They can directly process high-dimensional data, capture and remove noise from the data, and learn layer by layer through multiple hidden layers. MLPs can effectively capture the nonlinear features of the data. In addition, due to their low computational complexity, MLPs are more suitable for iterative training on large-scale data such as scRNA-seq data compared to other neural networks.

[0063] Then, the mean was fitted using three different fully connected layers ( ), discrete values ​​( ) and dropout probability ( ),as follows:

[0064] ,

[0065] in, This represents random Gaussian noise with a mean of 0 and a variance of 0.01, enhancing the model's denoising capability. Indicates the mean gene expression; Represents discrete parameters and measures the variability of gene expression levels; This represents the probability of the zero-inflation component. Each fully connected layer is activated by a different nonlinear function to enhance the model's generalization ability. Since the values ​​range from 0 to 1, the Sigmoid function is used for standardization, and the mean is... and discrete values Since the values ​​are non-negative, an exponential function is used for activation. To prevent overfitting, the mean is... All were normalized to the same extent as the original count size of each cell. This represents the learnable parameters.

[0066] The final loss function is defined as the Adaptive Zero-Inflated Negative Binomial Distribution (Adaptive-ZINB) function, which is based on adaptive weight adjustment of the zero-inflated ratio. It is primarily used to address the high zero-throughput problem in single-cell RNA sequencing data while considering the excessive dispersion of the negative binomial distribution. Compared to the traditional ZINB function, this loss function adaptively adjusts the weight ratio of the zero-inflated and negative binomial components in the dataset, avoiding overemphasis on the zero-inflated portion in low-throughput data and overemphasis on the negative binomial portion in high-throughput data. For example, in data from different tissue samples or different sequencing platforms (such as 10x Genomics vs. Drop-Seq), the zero-throughput ratio varies significantly. Traditional ZINB Loss cannot automatically adapt to different data distributions. The adaptive loss function can appropriately balance dropout and truly low-expressed genes, improve the true gene expression signal, enhance the detection capability of rare cell types, and avoid ignoring some truly low-expressed gene signals, thus preventing the loss of biological information. The formula is as follows:

[0067] (6)

[0068] in, These are observations in the cell matrix; Indicates an indicator function, when The value is 1 if the condition is met, otherwise it is 0. This represents the total number of elements in the dataset. We can define it as:

[0069] (7)

[0070] in, , For the defined adaptive weights, It is the weight of the zero-inflation portion, which increases with the proportion of zero. The weight of the negative binomial portion decreases proportionally with the zero value, from which we can deduce:

[0071] (8)

[0072] (9)

[0073] By introducing adaptive weights into the standard ZINB Loss formula, and adjusting the parameters adaptively, the model overfitting problem can be reduced, making it suitable for different sequencing platforms and improving the model's generalization ability.

[0074] GAE based on GCN:

[0075] Another strategy employs a graph autoencoder (GAE) to capture the topological relationships between cells and learns low-dimensional embeddings of graph nodes to reconstruct the graph's adjacency matrix. GAE optimizes graph representation learning through unsupervised learning. Unlike traditional GNNs that generate node embeddings based on features, this embodiment uses a graph convolutional network (GCN) as the encoder unit. Since scRNA-seq data typically contains a lot of noise and high sparsity, GCN can better capture and integrate the complex relationships between cells and utilize the inherent spatial information of graph data for feature extraction. GCN employs a local aggregation mechanism, which can effectively smooth noise in the data and enhance biologically relevant signals. The calculation formula for each GCN layer is as follows:

[0076] (10)

[0077] in, It represents a symmetric normalized adjacency matrix with self-loops added to ensure that information can propagate in the network; It is the first Layer node feature embedding represents a low-dimensional representation of a cell; It is the first The trainable weight matrix of the layer; activated by the hyperbolic tangent function (tanh), defined as follows:

[0078] (11)

[0079] The tanh function has nonlinear transformation capabilities, which can effectively alleviate the gradient vanishing problem, allowing deep GCNs to maintain strong expressive power.

[0080] In the decoding part, the dot product reconstruction method is used to calculate the dot product reconstructed adjacency matrix of the node embedding. The calculation formula is as follows:

[0081] (12)

[0082] in It is the reconstructed adjacency matrix with size . × , It uses the sigmoid activation function to ensure that the output value is between 0 and 1, matching the format of the input adjacency matrix. Finally, a mean squared error (MSE) loss function is established to measure the error between the reconstructed and original graphs.

[0083] (13)

[0084] in and The original graph and the reconstructed graph respectively represent the first two graphs in the first two graphs. row and number The elements of the column. By minimizing this loss function, the parameters of GAE are optimized so that the embedding representation better conforms to the topological structure between cells.

[0085] Compared to traditional graph neural network (GNN) methods for generating node embeddings, the GAE method in this embodiment can learn the relationships between cells more accurately and enhance the separability between cell populations while further reducing the impact of noise, which provides a more reliable basis for subsequent analysis.

[0086] Dynamic multi-head attention mechanism:

[0087] Dynamic Combination Multi-Head Attention (DCMHA) is an extension of the standard multi-head attention mechanism by introducing dynamic weights to adjust the contributions of different attention heads. Compared to traditional multi-head attention mechanisms, it introduces dynamic weights to dynamically adjust the contributions of each head based on the input data. This allows the model to adaptively handle different input scenarios. For example, it relies more on denoising features in noisy data, more on graph structure features in data with complex topologies, and performs better when dealing with nonlinear relationships in single-cell data, better capturing subtle differences and complex relationships in cell states. To fully combine the advantages of DAE and GAE and mitigate their respective shortcomings, the DCMHA combines denoising embedding with topological embedding, as shown in the following formula:

[0088] (14)

[0089] First, embed the topology from GAE. Denoising embedding from DAE Projected onto the query, key, and value spaces respectively. , , Both are learnable projection matrices. Next, an attention score is calculated using a dot product attention mechanism. The dimension of the key vector is used to scale the dot product result, prevent gradient explosion, and is normalized using the softmax function to convert the attention scores into a probability distribution. Then, the weights of each attention head are dynamically calculated based on the input data. , It is a learnable function used to generate weights. Finally, the output of each attention head is passed through dynamic weights. Weighted summation yields the final fusion result. Indicates the first The output of each attention head, It is a learnable output projection matrix. This is used here as a residual connection to preserve the information of the original denoised embedding, avoid dilution of key information during the fusion process, and alleviate the gradient vanishing problem.

[0090] Unsupervised deep embedding clustering:

[0091] An unsupervised algorithm is employed to optimize clustering until convergence. The algorithm first uses the Student's t-distribution as the kernel function to calculate the soft membership of each data point to the cluster center in the embedding space, evaluating the similarity between data points. Represents cells embedding point, Clustering The centers, which are derived from pre-trained embeddings by the Leiden algorithm. The Leiden algorithm obtains this information. Although the Leiden algorithm does not require pre-specifying the number of clusters... However, it is known It can help optimize initial conditions. For example, given... It can be used to set appropriate parameters, thereby improving the quality of clustering results. Furthermore, it is known that... It can serve as a benchmark for verifying the clustering results of the algorithm. Known... By comparing the results with the actual number of clusters obtained, the accuracy and effectiveness of the algorithm can be evaluated, and the algorithm can be further optimized. The kernel function formula is as follows:

[0092] (15)

[0093] in This represents the degrees of freedom of the student t-distribution, typically 1. The clustering is then further optimized by optimizing the distribution of the high-confidence cluster assignments. The distribution is defined as follows:

[0094] (16)

[0095] The clustering optimization process uses soft allocation and auxiliary allocation The Kullback-Leibler divergence loss is defined as follows:

[0096] (17)

[0097] The loss function is ultimately defined as follows:

[0098] (18)

[0099] (19)

[0100] This model was developed in Python 3.7, with hidden layer dimensions set to 256, 64, 16, 64, and 256 respectively. It includes two dynamically combined multi-head attention blocks, each consisting of eight heads. The DAE and GAE share a common linear reduction layer to improve training efficiency. The Adam optimizer is chosen with a learning rate of 0.001, and a gradient pruning strategy with a maximum L2 norm of 3 is used to prevent overfitting. During the fine-tuning phase, only clustering operations are performed; training is terminated early if the label change is less than 1 / 1000 of the total number of labels. and This is a parameter used to balance multi-objective optimization, with a default value of 0.1.

[0101] To evaluate the clustering performance of this model, two widely used clustering evaluation metrics were used: the adjusted RAND index (ARI) and normalized mutual information (NMI). Higher values ​​of these metrics indicate a better correspondence between predicted and true labels, suggesting better clustering performance. The true cell labels from the scRNA-seq data were set to... , , , … , The predicted cell label is set to = , , ,…, .

[0102] NMI (Nearest Migration) measures the similarity between predicted and actual results, ranging from 0 to 1. It combines mutual information and entropy. It addresses the sensitivity of MI to the number of clusters. The calculation formula is as follows:

[0103] (20)

[0104] in Used for calculation and Mutual information between them:

[0105] (twenty one)

[0106] and Representing tags respectively and Information entropy Indicates the total number of cells:

[0107] (twenty two)

[0108] (twenty three)

[0109] ARI is used to measure the overlap between predicted and actual clusters, and its range is from -1 to 1. The formula for ARI is:

[0110] (twenty four)

[0111] in express and The number of cells overlapping between them. Represents categories in the true classification The number of cells, Indicating clusters in the clustering results The number of cells.

[0112] result

[0113] Method comparison:

[0114] The scDAEC model was applied to cluster 13 real scRNA-seq datasets to obtain the final predicted labels, comprehensively evaluating the clustering performance of the scDAEC model. The clustering results were compared with several state-of-the-art deep learning methods, including scGAC, Graph-SCC, scGNN, and scDeepCluster, all of which have default parameters. These deep learning methods are based on autoencoders, graph neural networks, and contrastive learning. Through different clustering techniques and various types of single-cell deep clustering methods, the effectiveness of our proposed method was comprehensively demonstrated. Using the same preprocessing method, 2500 highly variable genes were selected from the original data as input for all methods. All clustering methods were repeated five times on each dataset, and the average value was taken. Furthermore, the adjusted RAND index (ARI) and normalized mutual information (NMI), two widely accepted clustering metrics, were used to evaluate the model's clustering performance.

[0115] Figure 2 and Figure 3The comparison results of 10 clustering methods on 13 datasets are shown. The 10 clustering methods are: Seurat, SIMLR, scDeepCluster, scGAC, scTAG, SC3, CIDR, scGNN, Graph-SCC, DESC, and scDAEC. The 13 datasets are: Adam, Baron_human, Baron_mouse, Chen, Colquitt, Elliott, Guo, Kakebeen, Muraro, Plasschaert, Quake_10x_Heart_and_Aorta, Quake_10x_Limb_Muscle, and Zilionis_mouse_lung.

[0116] As can be clearly seen from the figure, scDAEC outperforms the other 10 clustering methods on 13 datasets. The ARI score is improved by an average of 35.31%-60.73%, and the NMI score by an average of 37.55%-59.78%. Notably, compared to region-based methods such as Seurat, SIMLR, scDeepCluster, SC3, and CIDR, scDAEC does not require pre-determining the number of cluster centers, yet its clustering results are superior to these methods. Furthermore, compared to other methods that adaptively learn the number of clusters, this method outperforms other methods across various datasets. In specific experiments, it was found that when there are fewer cell clusters in the actual data, methods that require pre-determining the number of cluster centers perform better, while other methods prioritize dividing the dataset into more predicted cell clusters. Nevertheless, scDAEC still leads in overall performance. For the classic ZINB-based model scDeepCluster, the model in this method achieves better results because it fully utilizes the inter-cell relationships during the learning of cell features. SC3 underperforms deep learning-based methods, overestimates the number of clusters when using random matrices, and produces null values ​​on some large datasets. Overall, scDAEC has a significant advantage in clustering compared to other methods.

[0117] Clustering quality visualization verification:

[0118] To obtain a clear biological interpretation, the Muraro and Quake_10x_Limb_Muscle datasets were selected for comparison across four methods, as these two datasets are representative in terms of cell count. For each method, the clustering results were visualized in 2D space using UMAP. Figure 4 and Figure 5The comparison results of different methods on two datasets are shown, with different colors representing different cell types. The comparison chart clearly demonstrates the advantages of the embeddings learned by scDAEC. The clustering results of this method have clear boundaries, better separating different cell types. Because of the two clustering strategies, scDAEC can fully utilize the correlations between cells, learn useful information from the real dataset, and reduce the impact of noise, achieving better learning results. In contrast, other methods fail to cluster cells with the same labels together, resulting in more blurred boundaries. Graph-SCC and DESC have relatively clear boundaries, while scGNN and Scanpy mix some cell types with others. Although Graph-SCC and scGNN also use GAE to extract features, their results are not ideal because they separate representation learning and clustering into separate tasks. scDAEC not only considers the above methods, employing a dynamic combination of multi-head attention mechanisms combined with denoising and topological embedding to construct cell features, but also fully considers the biological information of low-expression cells, achieving optimal performance through comprehensive consideration.

[0119] To more directly demonstrate the ability of scDAEC to distinguish unknown cells, the common inner clustering evaluation method, the Silhouette coefficient, was used to assess clustering performance. The Silhouette coefficient, which combines cohesion and separation, ranges from -1 to 1, with values ​​close to 1 indicating good clustering and values ​​close to -1 indicating poor clustering. Experimental comparisons were conducted using latent features of location cells obtained from the deep model and the predicted results. Figure 6 As shown, this method exhibits the best performance, with the highest mean Silhouette coefficient, smaller interquartile range, and stable results, demonstrating strong cluster consistency. DESC also shows good performance, but its larger interquartile range makes it less stable than scDAEC. Graph-SCC and scGNN perform moderately, with lower Silhouette coefficients, potentially leading to cell confusion in some experiments. Scanpy clustering shows the worst results, highlighting the limitations of traditional methods on single-cell data. Overall, scDAEC not only demonstrates better clustering performance but also provides more stable results.

[0120] The Zillionis_mouse_lung dataset contains seven cell types: 'B cells', 'Basophils', 'MoMacDC', 'NK cells', 'Neutrophils', 'T cells', and 'Plasmacytoid Dendritic Cells'. The scDAEC proposed in this embodiment achieves high clustering accuracy on this dataset. To visually compare the performance of clustering methods, a Sankey plot is used to show the correspondence between the clustering results of each method and the true cell types, such as... Figure 7 Observations show that the Seurat and SIMLR methods incorrectly cluster the three major cell types into one category, Graph-SCC tends to group cells of the same type into multiple categories, and scDeepCluster and scGAC tend to mix some cell types with other cell types, resulting in a small number of cells in each category. In contrast, our method efficiently classifies each type.

[0121] scDAEC is beneficial for removing batch effects:

[0122] Batch effect removal is a crucial step in single-cell data analysis, helping to eliminate technical bias, restore authentic biological signals, and improve the reliability of downstream analysis tasks. Harmony, based on soft clustering and iterative optimization, maximizes the mixing of cells between batches while preserving biological variation. scVI, based on variational autoencoder (VAE) modeling, learns the biological variation and batch effects in the data. BBKNN balances batch effects in KNN, reconstructing an adjacency graph connecting similar cells from different batches. ComBat, based on a Bayesian framework, estimates batch effect parameters through normalization and linear models. scDAEC analyzed batch effect results from five scRNA-seq datasets and two simulated datasets, evaluating its batch effect removal performance from two core dimensions: batch effect removal capability and biological variation preservation capability. Figure 8 As shown, scDAEC performs best in batch effect removal while also maintaining a high level in biological variation preservation, demonstrating excellent integration capabilities and stronger stability and adaptability across different datasets. Part A of the figure shows the average score distribution of various integration methods across the two dimensions of batch effect and biological variation preservation. scDAEC is located in the upper right corner, indicating that it achieves a good balance between the two. Part B of the figure further shows the ranking of different methods on various datasets and tasks. scDAEC ranks highly in multiple tasks, demonstrating consistently excellent performance. The methods include scDAEC, scVI, Harmony, BBKNN, Scanorama, ComBat, MNN, Conos, and DESC.

[0123] In summary, the results show that scDAEC can effectively handle batch effects in data processing and is suitable for complex scRNA-seq data integration tasks involving multiple batches, multiple sources, and multiple modalities, providing a solid data foundation for downstream analysis.

[0124] scDAEC is beneficial for cell trajectory inference:

[0125] Reconstructing cell trajectories is a common task in scRNA-seq analysis, crucial for studying cell differentiation, cell cycle changes, and cellular responses to external stimuli. Typically, cell trajectory analysis first reduces the complexity of gene expression data to more effectively select important features, then constructs dynamic cell trajectory paths, and finally maps each cell to its corresponding position on this trajectory. Monocle3 is a widely used method for scRNA-seq data trajectory analysis, capable of generating corresponding cell trajectories from cellular data features. Therefore, in the experiment, the low-dimensional cell features obtained from scDAEC and the original feature matrix were input into Monocle3 to obtain cell time trajectories. The pseudo-temporal ranking score (POS) and Kendall's rank correlation score (Cor) were used as evaluation metrics to demonstrate the effectiveness of scDAEC dimensionality reduction. Kendall's rank correlation score assesses the consistency of ranking between two sets of observations, while POS reflects the relationship between the predicted temporal order and the actual event labels; both aim to obtain higher scores when compared. Using the Adam dataset, from... Figure 9 It can be seen that there is a gap between the cell trajectories reconstructed from the original data and the real event labels. In contrast, the pseudo-temporal trajectories of cells inferred from the data after dimensionality reduction by scDAEC have a high correspondence with the real event labels. In addition, scDAEC has higher POS and Kendall scores, which effectively proves that scDAEC is helpful in reconstructing cell trajectories and can effectively perform data dimensionality reduction.

[0126] Cluster center adaptive change:

[0127] scDAEC employs the Leiden algorithm to identify and establish initial cluster centers, analyzing cell distribution in the latent space. The model adaptively assigns labels to each cell, independent of soft assignment strategies that rely on a pre-defined number of groups. This process is independent of prior knowledge of subpopulation counting, providing robust support for exploring unknown or complex biological systems. Furthermore, because scDAEC is based on a dynamic multi-head attention mechanism, it adaptively processes data from different scenarios, better integrating denoised embeddings and topological embeddings. This results in cell representations that more closely approximate similar cells in the latent embeddings, distinguishing different cell states. In each iteration of the clustering phase, cells evaluate their similarity to each other through an attention mechanism, continuously adjusting the number and location of cluster centers to select the most suitable ones. Figure 10As shown, the Leiden algorithm initially tends to set more cluster centers. This strategy is overly granular at the beginning, but it provides the model with a more flexible way to explore the underlying structure of the data. As the model iterates through training, some of the initially set cluster centers are deemed redundant and discarded. This not only reduces the impact of noise but also ensures that the remaining cluster centers more accurately represent the true patterns of the data. Ultimately, dynamically adjusting the number and location of cluster centers effectively improves the overall clustering performance and guarantees the accuracy and reliability of the clustering results.

[0128] Model stability verification

[0129] scRNA-seq data is characterized by sparsity, which may lead to model instability on sparse datasets. Therefore, dropout is simulated by artificially setting the expression values ​​of some genes to zero. Dropout can cause some gene expression to be underestimated or completely absent, affecting data quality and leading to misclassification of cell types. In this example, eight scRNA-seq datasets were selected and artificially set to zero, and the experiment was repeated five times. The average ARI and NMI were used to evaluate clustering performance and verify the model's stability and robustness. In the specific experiments, 10%-50% of gene expression values ​​were randomly set to 0. Figure 11 As shown, 10%-20% artificial zeroing yielded better results than the baseline scenario, with no significant decline at dropout rates of 30%-50%. This indicates that the model can effectively extract features from sparse data for cell type inference, demonstrating its stability and robustness to scRNA-seq data. Dropout is considered a form of noise in most cases, and DAE's denoising capabilities play a significant role. In cases with more missing values, low-expressed genes may have a significant impact on classification. scDAEC balances dropout and truly low-expressed genes during denoising, thus achieving better results than the baseline model with lower artificial zeroing rates.

[0130] Ablation experiment:

[0131] scDAEC mainly consists of a feature extraction module and an information fusion module, with each component having a unique function. To evaluate the importance of each group in these two modules, various components were removed, and the performance of scDAEC was evaluated under the following conditions: no dynamic multi-head attention mechanism (no attention mechanism), no GCN-based GAE (no GAE), no DAE denoising (no DAE), and no adaptive ZINB (no balancing of dropout and truly low-expression genes). Experiments were conducted using eight real scRNA-seq datasets, and the results show the average ARI and NMI values. The eight real datasets are: Domingo-Gonzalez, Fincher, Gutierrez, Haber_Smart-seq2, Han, Ludwig, Quake_10x_Bone_Marrow, and Quake_10x_Mammary_Gland.

[0132] like Figure 12 and Figure 13 As shown, scDAEC exhibits excellent clustering performance. Replacing the attention layer with a multilayer perceptron results in a performance decrease. Removing the DAE leads to a decline in the model's denoising ability, affecting the scoring results of subsequent analyses, especially the NMI value, which drops significantly across various datasets, indicating that noise has a significant impact on the clustering results. The GCN-based GAE in the model is also crucial. Removing this component results in a lack of topological structure information, leading to a significant decrease in model performance, particularly the ARI value. This also reflects that clustering tasks should focus more on intracellular correlations, which is key to improving model efficiency. For the adaptive ZINB loss, replacing it with the standard ZINB loss resulted in a decrease in model performance. This indicates that balancing dropout and low-expression genes in this embodiment is effective. Some low-expression genes do indeed possess biological signals, which can significantly help with the clustering results, improve the accuracy of cell classification, and enhance cell type identification results.

[0133] In summary, the single-cell clustering method based on the scDAEC deep model proposed in this embodiment combines a novel adaptive ZINB loss function, balancing dropout and true low-expression genes, improving the true gene expression signal, avoiding ignoring true low-expression gene signals, and enhancing cell classification efficiency. scDAEC integrates two different scRNA-seq clustering strategies through a dynamic multi-head attention mechanism. Specifically, it utilizes a DAE (denoising autoencoder) based on an adaptive ZINB model to extract low-dimensional features, handle dropout events, and balance technical zeros. Simultaneously, a GAE (graph autoencoder) based on GCN uses neighboring cell information to guide dimensionality reduction and obtain topological embeddings. The information from these two different modalities is fused through a dynamic multi-head attention mechanism, dynamically adjusting the contribution of each head. This allows the model to adaptively handle different input scenarios, better capture subtle differences and complex relationships in cell states, obtain more clustering-friendly representations in the hidden embeddings, and enhance clustering results. Furthermore, because the model can dynamically balance dropout and true low-expression genes, it avoids errors caused by some technical zeros. scDAEC uses the Leiden algorithm to pre-initialize cluster centers, avoiding reliance on prior knowledge of population size. Experimental results show that scDAEC can effectively achieve single-cell data clustering, dimensionality reduction, and visualization. The model has good denoising capabilities, excellent fusion strategies, and stability and robustness.

[0134] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0135] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of the present invention. Furthermore, those skilled in the art will recognize that, based on the ideas of the present invention, there will be changes in the specific implementation methods and application scope. Therefore, the content of this specification should not be construed as a limitation of the present invention.

Claims

1. A deep clustering method for single-cell data based on a dual autoencoder, characterized in that, include: Obtain raw genetic data from single cells; Construct the scDAEC model; The scDAEC model includes a dynamically combined multi-head attention mechanism, a denoising autoencoder, and a graph autoencoder; the denoising autoencoder is constructed based on a zero-inflated negative binomial distribution loss adjusted by zero-inflated proportional adaptive weights; the graph autoencoder is constructed based on a graph convolutional network; The original gene data is input into the scDAEC model for clustering to obtain single-cell data clustering results; The step of inputting the original gene data into the scDAEC model for clustering processing to obtain single-cell data clustering results specifically includes: The raw gene data is preprocessed and sorted according to gene content to generate a gene expression matrix; the rows in the gene expression matrix represent cells and the columns represent genes. A directed graph is constructed based on the KNN algorithm and the gene expression matrix; The directed graph is input into the denoising autoencoder and the graph autoencoder respectively. In the denoising autoencoder, topological embeddings are generated by balancing dropout and real low-expression genes. In the graph autoencoder, the topological relationships between cells are captured to generate denoising embeddings. The topological embedding and the denoised embedding are combined using a dynamic multi-head attention mechanism to obtain single-cell data clustering results, specifically including: First, the topological embedding and the denoised embedding are projected onto the query, key, and value spaces, respectively. Second, the attention score is calculated using the dot product attention mechanism and normalized using the softmax function to convert the attention score into a probability distribution. Then, the weight of each attention head is dynamically calculated based on the input data. Finally, the outputs of each attention head are summed using dynamic weights to obtain the single-cell data clustering result.

2. The single-cell data deep clustering method based on dual autoencoders according to claim 1, characterized in that, The preprocessing includes quality control, standardization, logarithmic transformation, and normalization.

3. The single-cell data deep clustering method based on dual autoencoders according to claim 1, characterized in that, The specific processing procedure in the aforementioned graph autoencoder is as follows: A local aggregation mechanism is used to smooth noise in the data, and the following computations are performed in each layer of the graph convolutional network: ， in, It represents a symmetric normalized adjacency matrix, with self-loops added to ensure that information can propagate in the network; It is the first Layer node feature embedding represents a low-dimensional representation of a cell; It is the first The trainable weight matrix of the layer, where tanh represents the hyperbolic tangent function; In the decoding part, the dot product reconstruction method is used to calculate the dot product reconstructed adjacency matrix of the node embedding: ， in, It is the reconstructed adjacency matrix. It uses the Sigmoid activation function; finally, the mean squared error loss function is established to measure the error between the reconstructed image and the original image: ， in, and The original graph and the reconstructed graph respectively represent the first two graphs in the first two graphs. row and number The elements of the column are optimized by minimizing the mean squared error loss function to obtain a topological embedding that conforms to the intercellular relationships.

4. The single-cell data deep clustering method based on dual autoencoders according to claim 1, characterized in that, The zero-inflated negative binomial distribution loss, adjusted by the zero-inflated proportional adaptive weights, is: ， in, For adaptive weights, The weights for the zero-inflation portion, Indicates an indicator function, N This represents the total number of elements in the dataset. This represents the probability of the zero-inflation portion. This represents the negative binomial distribution probability mass function. Represents cells In genes The original count on, The parameter representing the mean of the negative binomial distribution. Represents discrete parameters.

5. An electronic device, characterized in that, The device includes a memory and a processor, the memory being used to store a computer program, and the processor running the computer program to cause the electronic device to perform the single-cell data deep clustering method based on a dual autoencoder according to any one of claims 1-4.

6. A computer-readable storage medium, characterized in that, It stores a computer program that, when executed by a processor, implements the single-cell data deep clustering method based on a dual autoencoder as described in any one of claims 1-4.

Images