APPLICATION AND COMPARISON OF DIFFERENT DEPTH LEARNING METHODS IN DATA FEATURE EXTRACTION

ABSTRACT

With the rapid development of single-cell sequencing technology, a vast amount of single-cell data is continuously being generated and accumulated. Identifying valuable information from this massive data has become a crucial task. Data mining, as an interdisciplinary field, is dedicated to analyzing large datasets to uncover patterns, key features, and other information, thereby extracting useful knowledge. Deep learning, a branch of machine learning, can achieve automatic feature extraction and representation learning from data, and thus has a wide range of application prospects in data mining..

The purpose of this thesis is to compare and analyze the application of different deep learning methods in bioinformatics data processing, and reveal the advantages and limitations of these methods in specific scenarios. In order to comprehensively evaluate the application effects of different deep learning models in the field of bioinformatics, this thesis makes an in-depth comparison and analysis of various models. These models include scDeepCluster, scGMAI, scGAC and so on, which have their own unique principles and functions and are suitable for dealing with different types of biological data. By comparing the performance of these models in processing single-cell RNA sequencing data and image data, it can be found that each model has its own unique advantages and applicable scenarios. For example, scDeepCluster effectively reduces and clusters high-dimensional data through a deep self-encoder, which is suitable for high-noise data; scGMAI uses generating confrontation networks to generate and modify data to make it outstanding in dealing with missing data and data enhancement; and scGAC uses graph convolution networks to capture the relationship network between cells, which is especially suitable for dealing with biological network data with complex structures. The difference between these models lies in their core algorithms and processing strategies. ScDeepCluster focuses on feature extraction and dimensionality reduction, scGMAI emphasizes data generation and modification, and scGAC focuses on relational modeling of structured data. Their respective advantages make them show specific application value and advantages in different bioinformatics data processing tasks.

Despite significant progress in applying deep learning to bioinformatics, improving model accuracy and generalization to handle more complex and diverse biological data remains a key future research direction. Innovations in optimizing algorithm design, reducing computational costs, and enhancing model interpretability will greatly advance the field of bioinformatics.

1.1 Background and significance of the research work
In the fields of bioinformatics and computational biology, with the rapid progress of high-throughput biotechnology, we can easily obtain massive biological data, including single-cell RNA sequencing data [1], spatial transcriptome data, and high-resolution biological image data. These data not only provide us with unprecedented insights, but also greatly promote our in-depth understanding of biological processes, providing powerful assistance for revealing disease mechanisms and promoting drug development.
In addition, the rise of deep learning technology has brought revolutionary changes to the field of data analysis. The excellent ability of deep learning in processing large-scale data sets makes it an indispensable and important tool in bioinformatics and computational biology [2]. The use of deep learning technology will help us analyze massive biological data more accurately and promote related research to a new stage.
In biotechnology, single-cell sequencing technology enables high-throughput analysis of gene expression in individual cells, revealing diversity and functional differences in cell populations. This revelation helps us more fully explore the specific roles of cells in biological processes and overcome the problem of limited cell numbers in samples to discover and study those rare cell types and states. This technology demonstrates its importance in several ways. In the study of cell development and differentiation processes, by tracking and analyzing the gene expression changes of single cells during development and differentiation, we can gain an in-depth understanding of the dynamic processes of cell development and differentiation, which is of great significance for understanding the complex mechanisms and regulatory networks of cell fate determination. In terms of personalized medicine and precision treatment, single-cell sequencing technology helps us better understand the genetic differences and disease risks between individuals through detailed analysis of individual cells, and provides personalized solutions for disease prevention, diagnosis and treatment. In addition, this technology can also help to mine new biological knowledge and drug targets [3], providing new ideas and methods for life science research and drug development.
The application of deep learning in single-cell data analysis further promotes the progress of these studies. The deep learning model realizes automatic identification and classification of cell types by learning and classifying features of single-cell data, improving the accuracy and efficiency of cell identification. In addition, deep learning methods can also mine potential patterns and regularities in single-cell data, reveal the differences and connections between different cell states and functions, and provide a more accurate and comprehensive basis for disease diagnosis and treatment [4]. In terms of gene regulatory network modeling [5], deep learning has also shown its unique advantages. By using deep learning models to learn the complex relationships and regulatory networks between genes, we can build more accurate and refined gene regulation models, thereby deeply understanding the regulatory mechanisms and signal transduction pathways of genes, and providing new theoretical support and treatment strategies for the occurrence and development of diseases.
In summary, single-cell sequencing technology combined with deep learning methods can not only deepen our understanding of cells and biological processes, but also bring new breakthroughs and hope for human health and disease treatment. Continuous development and innovation in this field will open up new paths for life science research and applications, making greater contributions to human health and well-being. Therefore, this research is not only of great theoretical significance but also has broad practical application potential. By systematically comparing and analyzing the application effects of different deep learning methods in data feature extraction, more scientific and efficient tools and methods can be provided for the analysis of single-cell sequencing data.
1.2 Research status
Research on deep learning methods in single-cell RNA sequencing (scRNA-seq), spatial transcriptome and other fields is booming and has achieved a series of significant progress.
In the field of scRNA-seq data analysis, deep learning technology has been widely used to improve analysis accuracy and efficiency. The DeepCCI deep learning framework [6] developed by the research team of Harbin Institute of Technology and Southern Medical University can effectively identify cell-cell interactions (CCI) from scRNA-seq data, highlighting the potential of deep learning in processing complex data sets to reveal cell-cell interactions. In addition, deep learning methods are also used to interpolate scRNA-seq data, such as non-negative matrix decomposition combined with Laplacian regularization and graph attention network, which further enriches the data processing methods.
In terms of spatial transcriptome data processing, deep learning technology also shows great strength. The STA series of tools [7] developed by Zhang Shihua's team at the Institute of Mathematics and Systems Science, Chinese Academy of Sciences, use artificial intelligence technology to promote the calculation and analysis of spatial transcriptome data and have made significant progress. At the same time, the development of deep learning-based spatial domain recognition technology (DeepST) [8] extracts image features from tissue morphology information through pre-trained deep neural network models, providing important support for spatial transcriptomic data mining.
1.3 Organizational structure of this article
Introduction: Introduce the research background, significance and current research status at home and abroad.
Theoretical basis: Explain the basic principles of deep learning technology, including scGMAI[8], scGAC[9], scDeepCluster[10], etc.
Data processing: Use code to calculate the zero-containing rate, cell number, and gene number of each data set, and evaluate various clustering performance
Comparison of methods: Using single-cell RNA sequencing data, analyze the feature extraction and clustering performance of different deep learning methods on single-cell data.
Result analysis: Compare the advantages and disadvantages of various methods and discuss their performance differences on different test data.
Discussion and Outlook: Based on the results of comparative analysis, the application prospects of deep learning technology in bioinformatics are discussed and possible improvement directions are proposed.
Conclusion: Summarize the main findings and prospects of this article.

Chapter 2 Principles and functions of different deep learning models
Deep learning methods have shown great potential and advantages in this field due to their excellent data processing and feature extraction capabilities. The following is a detailed analysis and comparison of the performance and advantages and disadvantages of several typical deep learning models: scGMAI[8], scGAC[9] and scDeepCluster[10] in processing single-cell RNA sequencing data and spatial transcriptome data.
2.1 scGMAI
scGMAI[8] (Single-Cell Generative Model for Artificial Intelligence) is an innovative method that uses generative adversarial networks (GANs) to process single-cell RNA sequencing data. GANs consist of two competing neural networks, a generator and a discriminator. The generator is responsible for generating virtual data, and the discriminator evaluates the authenticity of the data. Through this adversarial training mechanism, scGMAI is able to generate virtual data that is highly similar to real data, thus performing outstandingly in handling missing data and performing data enhancement. This method is particularly suitable for filling missing values ​​in gene expression matrices and can improve the robustness and generalization ability of the model through data augmentation.
The advantage of scGMAI [8] is that it can effectively simulate and correct single-cell data, improving data quality and integrity. For single-cell RNA sequencing data where technical noise or sparsity exists, scGMAI can significantly improve data accuracy.
However, this approach also presents some challenges. First, the training process of GANs is complex and requires high computing resources, which requires a delicate balance between the generator and the discriminator. Secondly, the stability problem of generative adversarial networks may lead to unstable quality of generated data, which puts forward higher requirements for practical applications. Despite this, scGMAI still has broad application prospects in data preprocessing and enhancement.
2.2 scGAC
scGAC[9] (Single-Cell Graph Attention Network) is a deep learning method based on graph convolutional networks (GCNs) that focuses on capturing the relationship network between cells. GCNs can process and analyze graph data with complex topological structures, and perform feature extraction on the graph structure through convolution operations. scGAC further introduces an attention mechanism, allowing the model to more effectively identify and focus on key inter-cell relationships, thereby improving the understanding of cell interactions and signaling pathways. When processing spatial transcriptome data, scGAC demonstrates its unique advantages. Spatial transcriptome data not only contains gene expression information, but also includes information on the spatial location of cells in tissues. scGAC uses this spatial information to construct a relationship network between cells, thereby revealing the spatial distribution and functional differences of cells in tissues. This has important implications for understanding the organizational structure and function of cells.
However, scGAC [9] has higher computational complexity and greater demand for computing resources, especially when processing large-scale graph data. In addition, the design and hyperparameter tuning process of graph convolutional networks are complex and require a large amount of experiments and verification to determine the optimal parameter settings. Nonetheless, scGAC has great potential and application prospects in revealing the complex relationships and spatial organization structures between cells.
2.3 scDeepCluster
scDeepCluster[10] is a clustering method based on Deep Autoencoder. It implements nonlinear data mapping and learns low-dimensional embedding representation of data by integrating the Zero-Inflated Negative Binomial (ZINB) model and autoencoder technology. This method introduces random Gaussian noise in the autoencoder part to enhance the low-dimensional representation quality of the data; in the decoder part, three fully connected layers are constructed to estimate the mean, dispersion and missing rate of the data. These three parameters jointly define the loss function of the ZINB model.
In addition, scDeepCluster [10] uses Kullback-Leibler (KL) divergence to measure the distribution difference between input data and reconstructed data, and defines a new loss function accordingly. Such a design not only helps optimize the parameters of the model, but also captures the intrinsic structure of the data more accurately. At the output end of the model, which is the low-dimensional space, the k-means algorithm is used for data clustering to further reveal potential patterns and structures in the data. Able to perform dimensionality reduction and clustering analysis on high-dimensional single-cell RNA sequencing data. Deep autoencoders compress high-dimensional data into low-dimensional representation through multiple nonlinear transformation layers to extract key features of the data. scDeepCluster takes advantage of this feature and performs well when processing high-noise data. It can effectively remove technical noise in the data through the denoising function of the autoencoder, thereby improving the accuracy and stability of clustering.
The main advantage of scDeepCluster [10] is its automatic feature extraction capability. Without pre-specifying features, the autoencoder can automatically learn and extract useful features. This gives scDeepCluster a significant advantage when processing high-dimensional and noisy single-cell RNA sequencing data. This method is particularly suitable for scenarios that require high-precision cluster analysis, such as the identification of subpopulation cell types and the discovery of novel cell populations.
However, the performance of scDeepCluster [10] is highly dependent on the structural design and training process of the autoencoder. The training time is long and the demand for computing resources is large. In addition, the black-box characteristics of deep autoencoders make it difficult to interpret their clustering results, which to a certain extent limits their application in actual biological research.
2.4 Summary
scGMAI[8], scGAC[9] and scDeepCluster[10] have their own characteristics in processing single-cell RNA sequencing data and spatial transcriptome data. scGMAI is outstanding in handling missing data and data enhancement, and is suitable for scenarios where data is sparse or technically noisy; scGAC has advantages in capturing inter-cell relationship networks and processing spatial transcriptome data, and can reveal the spatial distribution and functional differences of cells in tissues; while scDeepCluster has demonstrated strong capabilities in denoising and clustering analysis of high-dimensional data, and is suitable for high-precision clustering analysis of high-noise data.

Chapter 3 Extraction and comparison of different deep learning models in data features
3.1 Performance and differences of methods with different depth models on scRNA-seq data sets
3.1.1 Data
Below, 11 single-cell RNA sequencing (scRNA-seq) datasets are used to evaluate the performance of deep model-based imputation, visualization, and clustering methods. Data sources, cell counts, and clustering statistics for these datasets are summarized in Table 2.
Table 2-1 Summary of 11 data set benchmarks
Dataset Number of cells Number of genes Number of clusters Standardized type Cell resources Contains zero rate
Biase [11] 49 25737 3 FPKM Two- and four-cell mouse embryos 0.28
Yan [12] 90 20214 7 RPKM Human preimplantation embryos and embryonic stem cells 0.37
Goolam [13] 124 41480 5 UMI four-cell mouse embryo 0.68
Deng [14] 268 22457 10 RPKM mouse preimplantation embryo 0.32
Pollen [15] 301 23730 11 TPM Human 0.51
Kolodziejczyk [16] 704 38653 3 UMI mouse embryonic stem cells 0.48
Treutlein [17] 80 23271 5 FPKM Human lung epithelium 0.51
Ting [18] 149 29018 7 RPM Human pancreatic circulating tumor cells 0.52
Usoskin [19] 622 25334 11 RPM Human Neuron 0.78
Klein [20] 2717 24175 4 UMI human embryonic stem cells 0.69
Zeisel [21] 3005 19972 9 UMI mouse cortex 0.46

The data set of Biase[11] contains 9 fertilized eggs, 10 2-cell stage and 5 4-cell stage mouse embryos, with a total of 49 samples and 25737 genes. Gene expression values ​​were normalized by extracting fragments per thousand genes per million (FPKM) and applying the upper quartile normalization method.
The data set of Yan [12] included the transcriptomes of 90 individual human preimplantation embryos sequenced by Hiseq2000. Gene expression values ​​were normalized by applying the RPKM method, using 20,214 genes with expression values ​​greater than 0.1.
The data set of Goolam [13] [11] comes from all blastocyst cells of 28 embryos at the 2-cell, 4-cell and 8-cell stages, and 12 single cells of the 16-cell and 32-cell stages, with a total of 124 samples and 41,480 genes.
The data set of Deng [14] contains 268 single cells from fertilized eggs, late blastocysts and adult livers. Transcriptomes were normalized by using the RPKM method to generate read counts.
The dataset of Pollen [13] contains the transcriptomes of 301 single cells from 11 different populations, which were sequenced at low and high coverage. Gene expression values ​​for all samples were normalized by applying transcripts per million (TPM).
The data set of Kolodziejczyk [16] contains 250 serum cells, 295 2i cells and 159 a2i cells from mouse embryonic stem cells, respectively derived from three different culture conditions, with a total of 704 individual cells and 38,653 genes. And there are batch effects in different experiments under the three conditions.
The data set of Treutlein [17] contains 80 single-cell transcriptomes from mouse lung epithelial cells at four different stages, with a total of 23271 genes, and gene expression levels were quantified by the FPKM method.
The data set of Ting [18] contains 75 single-cell circulating tumor cells enriched in the blood of 5 mice, 12 single cells from one mouse embryonic fibroblast cell line, 16 single cells from the nb508 mouse pancreatic cancer cell line, 12 single mouse leukocytes, and 34 dilutions of 10 or 100 pg total RNA from mouse primary pancreatic tumors from 4 mice. Additionally, gene expression values ​​were normalized by reads per million (RPM).
The data set of Klein [20] includes 2717 transcriptome samples derived from mouse embryonic stem cells (mES) with 24175 genes. The gene expression values ​​of the dataset are presented as unique molecular identifier (UMI) filtered counts for each cell after matching and mapping without normalization.
The transcriptome of the Usoskin [19] dataset contains 622 dissociated single cells dissected from mouse lumbar dorsal root ganglia and distributed across a total of nine 96-well plates. Expression levels were quantified as read per million (RPM) values ​​for 25 334 genes.
The data set of Zeisel [21] includes unique molecular identifier (UMI) counts of 19 972 genes in 3005 individual cells from mouse somatosensory cortex and hippocampal CA1. Transcripts with gene expression values ​​less than 1 were considered not expressed and removed from further analysis of all data sets.
3.1.2 Data processing
Filter genes expressed in less than 10 cells, and normalize or log transform the data:
X=log2(X+1)
Contains zero rate calculation code:

import os

import pandas as pd

def calculate_zero_rate(file_path, encoding='ISO-8859-14'):

try:

data = pd.read_csv(file_path, sep='\t', encoding=encoding)

zero_rate_per_cell = (data == 0).sum(axis=1) / data.shape[1]

average_zero_rate = zero_rate_per_cell.mean()

return zero_rate_per_cell, average_zero_rate

except Exception as e:

print(f"Error processing file {file_path}: {e}")

return None, None

def process_all_files_in_directory(directory):

results = []

for filename in os.listdir(directory):

if filename.endswith('.tsv'):

file_path = os.path.join(directory, filename)

print(f"Processing file: {file_path}")

zero_rate_per_cell, average_zero_rate = calculate_zero_rate(file_path)

if zero_rate_per_cell is not None:

results.append({

'file': filename,

'average_zero_rate': average_zero_rate

})

output_file = os.path.splitext(filename)[0] + '_zero_rate.csv'

zero_rate_per_cell.to_csv(os.path.join(directory, output_file), header=['Zero_Rate'], index_label='Cell_Index')

print(f"Results saved to {output_file}")

Save summary of average zero rates to a CSV file

summary_df = pd.DataFrame(results)
summary_df.to_csv(os.path.join(directory, 'summary_zero_rates.csv'), index=False)
print("Summary of average zero rates saved to summary_zero_rates.csv")


Set the directory to the current directory

current_directory = os.getcwd()

process_all_files_in_directory(current_directory)

Cell number and gene number calculation code:
import pandas as pd
import os

def calculate_zero_rate_sparsity_and_dimensions(file_path):
try:
#Load CSV file
data = pd.read_csv(file_path)

    # 计算总元素数
    total_elements = data.size

# 计算含零的元素数
zero_count = (data == 0).sum().sum()

# 计算含零率
zero_rate = zero_count / total_elements

# 计算稀疏度，这里假设稀疏度与含零率相同
sparsity = zero_rate

# 获取细胞数（列数）和基因数（行数）
cells_count = data.shape[1]  # 列数
genes_count = data.shape[0]  # 行数

return zero_rate, sparsity, cells_count, genes_count

except Exception as e:
print(f"Error processing {file_path}: {e}")
return None, None, None, None


def main():
# Get all files in the current directory
files = os.listdir('.')

# 过滤出CSV文件
csv_files = [file for file in files if file.endswith('.csv')]

计算每个CSV文件的细胞数和基因数

results = {}
for csv_file in csv_files:
zero_rate, sparsity, cells_count, genes_count = calculate_zero_rate_sparsity_and_dimensions(csv_file)
if zero_rate is not None:
results[csv_file] = (zero_rate, sparsity, cells_count, genes_count)

输出结果

for file, metrics in results.items():
print(f"{file}: Zero Rate = {metrics[0]:.2f}, Sparsity = {metrics[1]:.2f}, Cells = {metrics[2]}, Genes = {metrics[3]}")


3.1.3 Performance of imputation method on scRNA-seq data set

Direct evaluation of imputation methods is difficult due to the challenge of defining the true value of a data set. Therefore, a single-cell RNA sequencing (scRNA-seq) data set was simulated by applying Splatter [22] to directly evaluate the performance of deep model-based denoising and interpolation algorithms. The simulated scRNA-seq data set contains 2000 cells and 2000 genes, divided into 6 groups. Three data recovery methods (scGMAI [8], scGAC [9], scDeepCluster [10]) were applied respectively to impute counts with dropout. By calculating the Pearson correlation coefficient (PCC) and root mean square error (RMSE) between imputed counts and true counts, it can be observed that scGAC significantly outperforms other methods in both PCC and RMSE. scDeepCluster is competitive on PCC but at the expense of larger RMSE. In contrast, scGMAI performs well on RMSE but relatively poorly on PCC. This shows that scGAC has excellent denoising and imputation performance, while other methods either ignore the correlation between genes or introduce imputation bias in the dataset.

Figure 2-1 Line chart of the interpolation performance of the three models under different inter-gene correlations and the same sample size

In terms of improving downstream analysis (such as clustering), scGAC [9] and scGMAI [8] perform well on almost all data sets, especially on the data set of Zeisel [21] and can significantly improve the clustering results. scGAC performs the most stably on all datasets, while scGMAI performs well on Treutlein [17] and Zeisel's datasets, demonstrating its excellent interpolation and denoising performance on some datasets, as well as instability on all datasets. But in fact, no method can guarantee that its interpolation process will definitely improve clustering performance, and each interpolation method has different preferences for data sets.

3.1.4 Evaluation performance of clustering methods on scRNA-seq data sets

The following will compare the clustering effects of four deep learning-based clustering algorithms on single-cell RNA sequencing (scRNA-seq) data sets. These algorithms all focus on cluster analysis and are implemented in Python. To evaluate their clustering performance, these methods were separately executed on the 11 datasets described above, using the default parameters provided by the methods and the same batch size for all four methods.

For scGNN [23], the default Louvain clustering method is used. Similarly, the adjusted Rand index (ARI), normalized mutual information (NMI) [24] and silhouette coefficient are used to evaluate the clustering results of these methods. Further, in order to compare with traditional clustering algorithms, the Louvain clustering algorithm and FEATS (a multifunctional tool for batch correction and downstream analysis) were directly executed on the 11 scRNA-seKolodziejczyk data sets applied in this article, and their clustering effects were evaluated. In terms of ARI and NMI indicators, scSemiCluster [25] shows advantages in almost all data sets, especially in the three data sets with larger cell numbers (et al., YanKlein et al., and Zeisel [19] etc.), achieved ARIs of 0.979, 0.908, and 0.75 and NMIs of 0.965, 0.88, and 0.728, respectively, indicating that the cluster labels assigned by scSemiCluster closely match the true situation of the cells.

Compared with the traditional clustering algorithm Louvain, scDeepCluster[10] also performs well, with significant improvements on several data sets, such as Goolam[13], Deng[14], Pollen[15], and Zeisel[21]. While scGNN [23] performs the most stably in terms of ARI, NMI and silhouette coefficient on almost all datasets, except the Treutlein [17] dataset. In terms of silhouette coefficients, scGNN outperforms scDeepCluster, scSemiCluster, DESC [26] and Louvain [27], which means that the separation between the clusters it assigns is greater and the compactness of cells within clusters is better. DESC also shows excellent stability on ARI and NMI indicators. Traditional clustering tools Louvain and FEATS also perform well on several scRNA-seq data sets, such as Biase[11], Yan[12], Pollen[15] and Klein[20], etc.

Furthermore, the scalability of three deep model-based clustering methods as well as two traditional clustering algorithms can be evaluated by evaluating the running time on data sets with different cell numbers. The results show that the clustering method based on deep models is more time-consuming than the classic clustering algorithm Louvain due to its complex structure and deep level. Among them, DESC demonstrates excellent and linear scalability, requiring only 36 seconds when the number of cells increases to 3000, while other methods require hundreds of seconds. Meanwhile, when the number of cells is less than 1000, the running time of scDeepCluster[10] and scSemiCluster[25] is tens of seconds. Overall, FEATS outperforms other clustering methods on the Usoskin [19] dataset.

Figure 2-2 scSemiCluster’s evaluation model based on the data set

However, their running time increases significantly when the number of cells exceeds 1000. Even with a cell number of 50, scGNN [23] takes hundreds of seconds due to its pruning cell graph process and graph autoencoder structure. As the number of cells increases, its running time slowly increases, indicating that scGNN also has good scalability for data sets with larger cell numbers. The running time of FEATS, another traditional clustering tool, increases exponentially with the number of cells in the data set, indicating that FEATS has relatively poor scalability, especially when dealing with data sets with high cell numbers. In general, for users with high requirements on clustering performance, scSemiCluster[25] and scDeepCluster[10] are more suitable, while users who consider the stability of the clustering algorithm are suitable to choose scGNN. When users need both good performance and short running time, they can choose DESC.

3.2.1 Advantages and Disadvantages of Graph Convolutional Neural Network (GCN)-scGAC

Graph convolutional neural network (GCN) [28] has excellent application results on single-cell RNA sequencing data and spatial transcriptome data. Especially in processing single-cell RNA sequencing data, the application of GCN has made significant progress.

scGAC[9] is a tool based on GCN[28], specifically used for interpolation and clustering of single-cell RNA sequencing data. It achieves faster execution than traditional GCN by simplifying the closed-loop architecture, and improves cell clustering performance by an average of 85.02% (as measured by the adjusted Rand index) across all datasets, and the median L1 error of interpolation is reduced by an average of 67.94%. This demonstrates that GCN is efficient and accurate in processing single-cell RNA sequencing data.

In addition, the scGAC [9] method shows superior performance in gene interpolation and cell clustering tasks by combining multi-layer perceptrons and graph neural networks (including attention networks). The study used indicators such as cosine similarity, median L1 distance and root mean square error to evaluate gene interpolation performance, and used criteria such as adjusted mutual information, normalized mutual information and completeness score to evaluate the effectiveness of different methods in cell clustering. Experimental results show that scGAC outperforms current single-cell data processing techniques in cell clustering and gene interpolation investigations.

By simplifying the closed-loop architecture, scGAC[9] achieves faster execution speed than traditional GCN[28], demonstrates excellent cell clustering and gene interpolation performance on multiple data sets, significantly improves clustering accuracy and reduces interpolation errors. It also demonstrates superior denoising and interpolation capabilities, and can effectively capture the complex relationships and interactions between cells. Compared with scGMAI and scDeepCluster, it has obvious advantages in capturing cell relationships and processing spatial transcriptome data.

Table 2-2 Data set derived from the deep model-based clustering methods of scGAC, scDeepCluster and scGMAI - ACC

Table 2-3 Data set derived from the deep model-based clustering methods of scGAC, scDeepCluster and scGMAI - ARI

Table 2-4 Data sets derived from the deep model-based clustering methods of scGAC, scDeepCluster and scGMAI - NMI

Table 2-5 Data set obtained by scGAC, scDeepCluster and scGMAI based on deep model clustering methods - Time

3.2.2 Advantages and Disadvantages of scGMAI

scGMAI [8] integrates an autoencoding network to effectively handle the common Dropout event in scRNA-Seq data, that is, many gene expressions are not detected and recorded as zero. By learning and reconstructing the data, the autoencoder improves the gene expression matrix and reduces the impact of dropout events on the analysis. In addition, scGMAI uses FastICA [29] for dimensionality reduction. This method performs well in identifying and highlighting independent features of the data, allowing scGMAI to capture the key features of scRNA-Seq data, providing better clustering accuracy than other methods such as PCA, t-SNE, UMAP, and ZIFA.

Clustering using Gaussian mixture models enables scGMAI [8] to efficiently cluster cells and identify cell types based on extracted features and dimensionality reduction data. The model automatically estimates the number of clusters, providing a robust framework for identifying cell populations in heterogeneous samples. By using t-SNE to visualize clustering results, cell distribution characteristics can be intuitively analyzed to help understand the spatial relationship and distribution between different cell types in complex scRNA-Seq data. Finally, scGMAI was tested on the scRNA-Seq dataset and showed better performance than other clustering methods on multiple evaluation metrics (e.g., NMI, ARI), indicating that scGMAI has good robustness under different datasets and conditions.

But multi-step processes including autoencoders, FastICA, and Gaussian mixture models can be computationally intensive, especially on very large datasets. The requirement for significant computing resources may limit the use of scGMAI in some environments. Although initializing the Gaussian mixture model using K-means++ helps optimize the clustering process [30], the overall performance of the algorithm used may still be sensitive to the initial conditions, which may affect the reproducibility and stability of the clustering results in some cases.

Although the autoencoder effectively handles Dropout events, the model's specific response strategies for different types of Dropout (random vs. systematic) are not detailed. If the characteristics of a dropout change significantly between datasets, this may impact performance. Furthermore, the effects of FastICA and Gaussian mixture models depend heavily on parameter selection, such as the number of components in FastICA and the initial conditions of the Gaussian model. Improper parameter settings may lead to suboptimal clustering results.

The efficient execution and accuracy of scGAC [9] make it superior in cell clustering and gene imputation tasks, but it also faces problems of high computational complexity and dependence on data quality. Therefore, scGAC has significant advantages in processing cell relationships and spatial transcriptome data.

3.2.3 scDeepCluster advantages and disadvantages

scDeepCluster [10] shows its superiority in processing different dropout rates and clustering signal strengths on simulated data sets. When the dropout rate is 30%, its NMI value is close to 1, showing excellent clustering performance, and even when the dropout rate increases to 30%, the performance remains at a high level of NMI of about 0.9, far exceeding other methods. In addition, when processing real scRNA-seq data sets, scDeepCluster also demonstrated high NMI, accuracy (CA) and adjusted Rand index (ARI), proving its effectiveness and accuracy in practical applications.

scDeepCluster[10] shows excellent clustering performance on multiple test data sets. In simulation evaluations under different dropout rates and clustering signal strengths, scDeepCluster consistently outperforms other methods (such as DCA+k-means, MPSSC, SIMLR, CIDR, PCA+k-means, scvis+k-means, DEC). Especially in the case of high dropout rate, the stability and superiority of its performance are particularly significant, which proves the method's effective processing ability of dropout events.

Moreover, scDeepCluster [10] can process data containing spatial location information and can adapt to data under different biological conditions [31], including different cell types and states. This method may use deep learning technology to process high-dimensional, high sparsity and high-noise scRNA-seq data, which helps to extract more representative features, thereby improving the quality of clustering.