scVAG: Unified single-cell clustering via variational-autoencoder integration with Graph Attention Autoencoder Authors: Seyedpouria Laghaee, Morteza Eskandarian, Mohammadamin Fereidoon, Somayyeh Koohi Published: 2024 ABSTRACT AbstractSingle-cell RNA sequencing (scRNA-seq) enables high-resolution transcriptional profiling of cell heterogeneity. However, analyzing this noisy, high-dimensional matrix remains challenging. We present scVAG, an integrated deep learning framework combining Variational-Autoencoder (VAE) and Graph Attention Autoencoder (GATE) for enhanced single-cell clustering. Building upon scGAC, our approach replaces its restrictive linear principal component analysis (PCA) with nonlinear dimensionality reduction better suited for scRNA-seq data. Specifically, we integrate VAE and GATE to enable more flexible latent space encoding. Extensive experiments on 20 datasets demonstrate scVAG's superior performance over previous state-of-the-art methods including scGAC, SCEA, SC3, Seurat, scGNN, scASGC, DESC, NIC, scLDS2, DRJCC, sLMIC, and jSRC. On average, scVAG improves clustering accuracy by 5 percent in ARI and 4 percent in NMI parameters. Visualizations highlight scVAG's capacity to recover interpretable biological structures. Our VAE-GATE pipeline extracts intricate expression patterns into compact representations that precisely delineate cell subpopulations consistent with ground truth labels. Overall, scVAG establishes a robust architecture for elucidating cell taxonomies from noisy transcriptomic inputs.Keywords: Single-cell RNA sequencing, Variational-autoencoder, Graph attention autoencoder, Clustering, Dimensionality reduction Introduction Cells are the fundamental units governing organism development through intricate gene interaction networks, giving rise to diverse, specialized cell types. While cellular heterogeneity enables population adaptation, it can also lead to disease if abnormal changes occur. Developing targeted treatments requires comprehensively decoding cell functionality across contexts [1]. By using RNA sequencing technology, it becomes possible to profile expression patterns of regulatory genes that govern cell behaviour. While traditional sequencing (Bulk RNA sequencing) measures average gene expression and ignores cell-to-cell variability, single-cell RNA sequencing changes cellular biology. Single-cell profiling has revolutionized our understanding of tissue formation, disease progression, and clinical response by unveiling novel cell types, trajectories, and communication networks. By uniquely barcoding mRNA from thousands of cells and capturing sequences, expression levels of more than ten thousand genes are quantified simultaneously across a cell population [2]. This provides unprecedented resolution into cell heterogeneity that is obscured in traditional bulk sequencing approaches [3]. However, realizing the full potential of scRNA-seq remains challenging due to the high dimensionality and technical issues inherent to the data. A major characteristic is the abundance of zero counts, stemming from either biological or technical factors. Biological zeros result from genes not expressed in a cell, while technical zeros occur when low abundance transcripts are lost during library processing and amplification. Discriminating between these zero types poses analytical difficulties [4]. Thus, appropriate computational approaches are critical to handle the sparsity and high dimensionality of scRNA-seq data [5]. Clustering analysis is crucial for understanding intricate cellular patterns. Moreover, clustering often forms the foundation for downstream differential expression testing, trajectory reconstruction, and cell annotation. It is important to note that, traditional clustering methods face limitations in handling noisy, high-dimensional data. They make assumptions about data distributions and require manual tuning of parameters like cluster numbers, risking suboptimal, biased results. In contrast, deep learning and graph-based methods provide compelling solutions by learning representations directly from gene expression patterns without restrictive assumptions. Complex neural networks and graph structures can intrinsically model nonlinear relationships in high-dimensional data, sensitively revealing rare subpopulations and transitional states indistinguishable to prior techniques [6]. To comprehensively evaluate our model, we compared it against many state-of-the-art single-cell gene expression clustering methods. Several state-of-the-art single-cell gene expression clustering methods have been introduced in recent years. Seurat [7] is widely used for integrating various data types and enabling both clustering and single-cell data inference. DESC [8], an autoencoder-based model, maps scRNA-seq data to a lower-dimensional space for clustering. SC3 [9] integrates multiple techniques and utilizes consensus clustering for robust results. The scASGC [10] model employs adaptive graph convolution with an attention mechanism for cell representation. scGNN [11] uses Graph Neural Networks for embedding, imputation, and clustering but is limited by ineffective graph denoising. scGAC [12] and SCEA [13] leverage Graph Attention mechanisms and use Network Enhancement [14] to denoise graphs. However, scGAC relies on Principal Component Analysis (PCA) for initial dimensionality reduction, which may not be optimal for complex high-dimensional data. SCEA uses a Multi-Layer Perceptron (MLP)-based encoder but lacks mechanisms to regulate learning effectively. Our model was also evaluated against several MATLAB-based methods, including sLMIC [15], NIC [16], jSRC [17], and DRjCC [18]. sLMIC integrates multi-omics data through self-representation learning, extracting shared and specific features. NIC clusters single-cell data using adaptive graph learning for integrated analysis. jSRC combines dimension reduction and clustering to enhance interpretability and scalability. DRjCC jointly learns dimension reduction and clustering, integrating projected matrix decomposition with non-negative matrix factorization. Additionally, scLDS2 [19] uses adversarial learning to effectively identify rare cell types. These comparisons demonstrate the robustness and versatility of our approach, tested against diverse baseline models with different clustering strategies. In our study, we present scVAG, a novel approach integrating Variational-Autoencoder (VAE) [20] and Graph Attention Autoencoder (GATE) [21] to enhance single-cell analysis. Building upon the scGAC framework, scVAG incorporates diverse Autoencoders and replaces PCA with a nonlinear dimensionality reduction structure. This improves downstream clustering accuracy. By leveraging the strengths of VAE in capturing latent representations of single-cell data and GATE in modeling complex relationships within graph-structured data, scVAG aims to capitalize on the advantages of both structures. The enhancements are attributed to the integration of VAE and GATE components, enabling more precise representation learning and capturing intricate relationships in single-cell gene expression patterns. The performance of scVAG was assessed on 19 diverse real and simulated datasets, using three widely-adopted clustering evaluation metrics. ScVAG outperformed state-of-the-art methods, including graph attention-based approaches like SCEA and scGAC. Results Overview of scVAG workflow A robust analysis of single-cell RNA sequencing data necessitates a well-defined pipeline including data preprocessing for cleanliness and normalization of data, dimensionality reduction for creating a meaningful low-dimensional representation, and clustering for identifying boundaries between cell subpopulations. This integrated approach ensures effective visualization and discovery of cellular heterogeneity in the high-dimensional gene expression profiles [22]. Building upon this need, our method, scVAG, implements a comprehensive four-step process to ensure robust and accurate clustering outcomes in single-cell gene expression analysis. As shown in Fig. 1, scVAG's methodology comprises four stages: I) data preprocessing and graph denoising; II) VAE-based dimension reduction; III) embedding generation via Graph Attention Autoencoder; and IV) self-optimized clustering for cell identification. This multi-step approach enables scVAG to effectively reduce high-dimensional single-cell data into a meaningful low-dimensional representation for revealing distinct cell characteristics, delineating boundaries, and accurately clustering cell subpopulations. Overview of the scVAG methodology. The workflow is divided into four main stages: I. Data preprocessing, graph construction, and graph denoising using Network Enhancement (NE); II. Variational Autoencoder (VAE) for dimensionality reduction of preprocessed data; III. Graph Attention Autoencoder (GATE) combining VAE results and denoised graph data to generate new embeddings using attention mechanisms; IV. Self-optimized clustering for final cell type identification. This pipeline integrates single-cell gene expression data processing, graph-based analysis, and deep learning techniques to achieve accurate and interpretable cell clustering results. Data preprocessing is a critical first step for reliable single-cell clustering. Careful data preprocessing, including filtering low-quality cells and genes, can reduce technical noise in scRNA-seq measurements. As demonstrated by Scanpy [23], tailored preprocessing enhances the quality of gene expression matrix for subsequent clustering. The cell-cell association graph is then constructed, revealing topological connections between similar cells and easing boundary identification. Subsequently, denoising methods are applied to refine the graph, enhancing its reliability and suitability for further analysis. Building upon the denoised graph, scVAG employ an integrated pipeline that combines a Variational-Autoencoder (VAE) with a Graph Attention Autoencoder (GATE) for nonlinear dimensionality reduction. By using these two different structures in dimension reduction part, this approach leverages both global data distributions and local graph structures to effectively reveal patterns among cell subpopulations. The decision not to use VAE or GATE independently stems from practical considerations. GATE's performance tends to decrease with a large number of features, and it requires substantial time to run on extensive gene expression profiles. This is primarily due to the calculation of attention weights, which scales with the number of input features. On the other hand, using VAE alone would miss out on graph information, which is essential for capturing intercellular relationships. Therefore, scVAG integrates VAE and GATE components to take advantage of their complementary strengths. We performed ablation experiments to assess the individual contributions of VAE and GATE components. According to Fig. 2, the best result was consistently achieved when the model used a combination of VAE and GATE in the pipeline across various datasets. This combination not only improves clustering accuracy but also enhances the model's scalability to larger datasets. Results of the ablation experiment, comparing model performance using VAE, GATE, and their combination (ARI (A) and NMI (B)). Compared to using VAE or GATE alone, the combined approach exhibits superior performance. In addition, we observed significant time efficiency improvements in our development process. For instance, using the Chung dataset (approximately 500 cells) on a Google Colab GPU, the full scVAG model completed its run in about 15 min, whereas GATE alone on the full feature set required approximately 319 min. This example illustrates the potential computational benefits of our combined approach. These observations underscore the synergistic benefits of combining VAE and GATE in the scVAG pipeline, demonstrating improvements in both performance and practical applicability. Finally, a self-optimizing clustering method is used for more concrete clustering. A standard algorithm such as K-means may produce suboptimal results due to random center initialization. To address this, scVAG integrates a self-optimizing mechanism into K-means. In the methods section, each component of the pipeline will be discussed in greater detail. Evaluation metrics for clustering In the assessment of our proposed method, we utilized three widely recognized metrics: Adjusted Rand Index (ARI) [24], Normalized Mutual Information (NMI) [25], and Cosine Similarity (Cs) [26]. These quantify the similarity between predicted and true cluster labels, providing a comprehensive evaluation. As shown in Eq. (1), the Adjusted Rand Index is calculated using the following formula. In the Adjusted Rand Index formula, n(i.j) represents the number of cells assigned to cluster i based on the model's predicted labels and to cluster j based on the true labels. a(i) denotes the total count of cells predicted to be in cluster i. Similarly, a(j) signifies the total count of cells actually belonging to cluster j based on true labels. Finally, n corresponds to the total number of clusters. The Adjusted Rand Index accounts for chance in cluster assignments, making it suitable for random labeling scenarios. The next metric, Normalized Mutual Information is calculated using the formula shown in Eq. (2). where, MI(X.Y) represents the mutual information between the assigned clustering X and true labels Y, H(X) denotes the information entropy for prediction X, and H(Y) is the information entropy of the true labels. Normalized Mutual Information offers a normalized perspective of mutual information values between assigned and true labels, facilitating interpretation. Both ARI and NMI offer valuable insights into the clustering accuracy, particularly in scenarios where direct label matching may be impractical due to the nature of clustering algorithms. In addition to the ARI and NMI, we utilized Cosine Similarity as another evaluation parameter for assessing the efficacy of our proposed clustering method. Cosine Similarity measures the cosine of the angle between two vectors and is commonly used to quantify the similarity between two clusters. The Cosine Similarity between two vectors X and Y is calculated as follows: where X⋅Y represents the dot product of vectors X and Y, and ∥X∥ and ∥Y∥ denote the Euclidean norms of vectors X and Y, respectively. In the context of clustering evaluation, Cosine Similarity offers insights into the alignment between the cluster assignments generated by our method and the true labels of the data. By quantifying the similarity of cluster assignments to the ground truth, Cosine Similarity provides an additional perspective on the clustering performance beyond what ARI and NMI offer. Overview of datasets We used 20 challenging real and simulated single-cell datasets from diverse biological contexts in order to provide a comprehensive evaluation of scVAG. The real single-cell datasets used in the project range from human to mouse samples, as summarized in Table 1. As listed in this table, the gene counts range from 16,412 to 36,601 and the cell numbers range from 49 to 5483. Our dataset package covers data with varying cell counts - from small datasets like Biase with 49 cells to large datasets like PBMC with 5483 cells. Table 1Summary of real single-cell datasets used to evaluate scVAG.Table 1Dataset NameType# of clusters# of cells# of genessequencing platformDate of releasedAccess linkKlein [27]Homo sapiens4271724021Illumina HiSeq2015GSE65525Romanov [28]Mus musculus7286318496Illumina HiSeq2017GSE74672Chung [1]Homo sapiens551527420Illumina HiSeq2017GSE75688PBMC [29]Homo sapiens842201641210X201710x GenomicsBjörklund [30]Homo sapiens464726,087Smart-Seq22016GSE70580Petropolus [31]Homo sapiens5151821627Illumina HiSeq2016E-MTAB-3929Brain_Mouse [32]Mus musculus550119465spaceranger-2.0.1202310x GenomicsNeuron_5K [33]Homo sapiens11548332286cellranger-6.0.0202010x GenomicsBiase [34]Mus musculus34921489Illumina HiSeq2014GSE57249Human_TBNK_Antibody [35]Homo sapiens589236601cellranger-6.0.0202110x GenomicsAdult_Mouse_Heart_5k [36]Mus musculus10322019925cellranger-7.0.0202210x GenomicsHEK293T & NIH3T3(HEK-NIH) [37]Homo sapiens & Mus musculus3592372302cellranger-8.0.0202410x Genomics Summary of real single-cell datasets used to evaluate scVAG. In addition, we generated 8 synthetic datasets with splatter [38] that covered a variety of situations, including varying numbers of clusters and imbalanced cluster distributions, to evaluate the model's performance under varying data characteristics. Table 2 provides detailed information of these datasets. Finally, it is worth noting that the datasets used in this project were obtained from public data repositories and original publications. The cell type labels for each dataset were derived from the metadata provided with the downloaded data files. Table 2Summary of synthesis single-cell datasets used to evaluate scVAG.Table 2Datasets #distributions# of cells# of genes# of clustersData1Unbalanced100020004Data2Balanced100020004Data3Balanced104020008Data4Unbalanced100020008Data5Unbalanced200020008Data6Balanced208020008Data7Unbalanced2240200016Data8Balanced2240200016 Summary of synthesis single-cell datasets used to evaluate scVAG. Data analysis In this section, we delve into a comprehensive analysis of scVAG's performance, emphasizing its clustering accuracy compared to twelve other baseline methods (SCEA, scGAC, scGNN, Seurat, DESC, SC3, scASGC, DRJCC, NIC, scLDS2, jSRC, and sLMIC). The primary metrics under consideration are the ARI and NMI, offering a well-rounded evaluation across diverse single-cell datasets. ARI and NMI comparisons between scVAG and other baselines are shown in Fig. 3. As illustrated, scVAG consistently outperforms the selected baselines across both ARI and NMI metrics. According to the average subplot of ARI and NMI, scVAG is capable of capturing meaningful subpopulations of cells in various biological conditions. To evaluate scVAG's robustness, we present two additional visualizations in Fig. 4. The first is a diagram depicting clustering accuracy vs. dataset size, with color brightness indicating performance for both ARI and NMI metrics. The second uses a heatmap showing accuracy values within each cell, with color brightness again representing clustering accuracy. Both visualizations demonstrate scVAG's robust performance. Comparison of clustering performance for scVAG versus baseline methods on 12 single-cell datasets. (A) Adjusted Rand Index (ARI) scores across all datasets. (B) Normalized Mutual Information (NMI) scores across all datasets. (C) Average ARI and NMI scores across datasets. ScVAG consistently achieves higher ARI and NMI scores compared to other methods, demonstrating its superior clustering performance in single-cell analysis. Comprehensive Robustness Evaluation of scVAG Across Two Metrics. Top row (Panels A and B) presents the Adjusted Rand Index (ARI) evaluation, with dot plots (A) and heatmap (B) demonstrating scVAG's robust performance across datasets compared to baselines. The bottom row (Panels C and D) illustrates the Normalized Mutual Information (NMI) assessment, mirroring the structure of the top row with corresponding dot plots (C) and heatmap (D). Consistent bright colors across all panels highlight the reliable clustering capabilities of scVAG, surpassing other methods on average. Furthermore, to evaluate the noise-handling capability and effectiveness of our proposed model under challenging conditions, we conducted experiments using 8 synthetically generated datasets exhibiting varying degrees of class number, cluster imbalance, and noise exposure. Fig. 5 shows how our model consistently achieved better results on simulated datasets. Comparison of Cosine Similarity scores for scVAG across 8 simulated datasets. Overall, scVAG demonstrates effective performance in measuring cell similarity across diverse simulated datasets. Fig. 6 utilizes t-distributed Stochastic Neighbor Embedding (t-SNE) to visualize scVAG's clustering findings on seven datasets: Chung, Human_TBNK_Antibody, HEK-NIH, Klein, PBMC, Petroplus, and Romanov. Visualizations comparing scVAG against baseline models on seven diverse single-cell datasets. The datasets are presented in rows from top to bottom: Chung, HEK-NIH, HTA(Human_TBNK_Antibody), Klein, PBMC, Petropolus, and Romanov. Each row presents results from 14 different methods alongside True labels for direct comparison of clustering performance. Across all datasets, scVAG's clusters consistently show strong agreement with the true cell type labels, accurately capturing intricate patterns in the single-cell data. This comprehensive comparison demonstrates scVAG's advanced ability for accurate and interpretable clustering across a wide range of biological datasets, outperforming current state-of-the-art methods in various single-cell contexts. The visualizations demonstrate scVAG's interpretability in accurately identifying meaningful biological structures. In the Chung data, scVAG distinctly separates B cells and T cells, superior to other methods. For the Human_TBNK_Antibody data, scVAG's clusters strongly correspond to the true labels, surpassing even current state-of-the-art baselines. These interpretations highlight scVAG's ability to uncover insightful single-cell structures by distinguishing underlying subpopulations in both datasets more accurately than other approaches. Statistical analysis To demonstrate the statistical significance of scVAG's performance improvements over state-of-the-art methods, we conducted rigorous statistical analyses. For each dataset and each evaluation metric (ARI and NMI), we ran each method 5 times to account for potential variability in performance. We then employed two-tailed paired t-tests to compare the mean performance of scVAG against each baseline method across all datasets. The null hypothesis for each test stated that there was no significant difference in the mean performance (measured by ARI and NMI) between scVAG and the compared method. By using multiple runs, we ensure a more robust statistical comparison that accounts for the inherent variability in the clustering algorithms. Fig. 7 presents 24 subplots illustrating the variations in ARI and NMI values for scVAG, SCEA, SC3, scGAC, scGNN, Seurat, DESC, and scASGC across all real datasets. The p-values resulting from these t-tests, based on the 5 runs for each method-dataset combination, are indicated on each plot. Statistical analysis of scVAG performance compared to state-of-the-art methods (SCEA, SC3, scGAC, scGNN, Seurat, DESC, and scASGC) across real datasets. (A) includes Adjusted Rand Index (ARI) subplots and (B) Normalized Mutual Information (NMI). Stars indicate the significance levels of the p-values obtained from t-tests comparing scVAG's performance against other methods, with the null hypothesis stating no average improvement by scVAG. We used the following significance levels. •∗ indicates p < 0.05 (significant) ∗ indicates p < 0.05 (significant) •∗∗ indicates p < 0.01 (highly significant) ∗∗ indicates p < 0.01 (highly significant) •∗∗∗ indicates p < 0.001 (very highly significant) ∗∗∗ indicates p < 0.001 (very highly significant) •∗∗∗∗ indicates p < 0.0001 (extremely highly significant) ∗∗∗∗ indicates p < 0.0001 (extremely highly significant) •ns indicates p ≥ 0.05 (not significant) ns indicates p ≥ 0.05 (not significant) For the p-value test, a smaller value indicates stronger evidence against the null hypothesis, suggesting a more significant difference between the methods. As an example, in the subplot for ARI on the Adult_Mouse_Heart_5k dataset (first row, first column), statistical tests indicate that scVAG outperformed all other methods (SCEA, SC3, scGAC, scGNN, DESC, Seurat, and scASGC) with p < 0.0001, denoted by ∗∗∗∗. This suggests an extremely highly significant improvement across the multiple runs. Across all the test cases, we passed the statistical test in over 90 % of the situations. Analysis of Marker genes To demonstrate the practicality of our model, we used the clustering results of the scVAG to conduct differential expression gene analyses on breast cancer data (GSE75688) [39]. We utilized the DESeq2 package to perform differential expression analysis. DESeq2 applies a negative binomial distribution to model the read counts and employs statistical techniques to identify genes with significant expression changes between the tumor and healthy conditions. The results were ordered by adjusted p-value, and genes with an adjusted p-value less than 0.05 and a positive log2 fold change were selected as differentially expressed. As listed in Table 3, CALML5, CXADR, ID4, TSPAN8, CLDN4, NDST4, KLK5, KRT81 are among the differentially expressed genes with adjusted p-values less than 1e-26. Table 3Differentially Expressed Genes Identified by scVAG Clustering Results in Breast Cancer Data (GSE75688), Selected Using DESeq2 Analysis with Adjusted p-Values < 1e-26 and Positive Log2 Fold Change.Table 3ENS IdGene Symbollog2FoldChangep-valueadjusted p-valueENSG00000178372CALML57.7620307654.76E-521.19E-48ENSG00000154639CXADR5.0622557812.29E-413.82E-38ENSG00000138653NDST49.4699219711.33E-331.43E-30ENSG00000167754KLK511.202035721.77E-321.78E-29ENSG00000205426KRT817.7519811861.63E-311.54E-28ENSG00000172201ID45.9728730643.80E-313.17E-28ENSG00000127324TSPAN87.318105051.57E-291.12E-26ENSG00000189143CLDN42.9731654019.91E-296.20E-26 Differentially Expressed Genes Identified by scVAG Clustering Results in Breast Cancer Data (GSE75688), Selected Using DESeq2 Analysis with Adjusted p-Values < 1e-26 and Positive Log2 Fold Change. There is evidence that all of these differentially expressed genes contribute to the development and progression of breast cancer. CALML5, a member of the calmodulin-like protein family, has been shown to have significant activity in primary breast cancers [40]. Loss of CXADR expression has been associated with epithelial-mesenchymal plasticity, a key process in breast cancer [41]. ID4, an inhibitor of differentiation proteins, is linked to a stem-like phenotype and poor prognosis in basal-type breast cancer [42]. Tspan8 is expressed in breast cancer and regulates E‐cadherin/catenin signaling and metastasis accompanied by increased circulating extracellular vesicles [43]. CLDN4, a tight junction protein, enhances cell proliferation, migration, tumor growth, and metabolic reprogramming in breast cancer cells [44]. NDST4, a gene encoding an enzyme involved in heparan sulfate biosynthesis, is frequently mutated in breast cancer [45]. KLK5 and KRT81 have been identified as potential serological biomarkers and indicators of aggressive pathological features and poor outcomes in breast cancer [46,47]. The significant enrichment of these genes among the DEGs highlights their importance in the molecular mechanisms driving breast cancer progression and metastasis, further validating the significance of our model. Discussion In this paper, we introduce scVAG, a novel deep learning framework for single-cell RNA-seq clustering analysis that integrates a Variational-Autoencoder and Graph Attention Autoencoder. Our approach outperforms the current graph clustering technique (scGAC) by replacing restrictive linear assumptions with flexible nonlinear dimensionality reduction specifically adapted to single-cell transcriptional profiles. Extensive comparative evaluations across 20 datasets spanning diverse biological contexts demonstrate scVAG's superior performance. Additionally, scVAG surpasses graph neural network techniques such as SCEA, scGAC, scASGC, and scGNN on all real datasets tested. This demonstrates its superiority as the top graph-based method for single-cell clustering. On average, scVAG achieves an Adjusted Rand Index of 0.693 and a Normalized Mutual Information of 0.733, outperforming previous methods including SC3, Seurat, scGAC, SCEA, scASGC, scGNN, and DESC. Based on statistical tests (t-tests), our model significantly outperformed other baseline models in more than 90 % of cases. Additionally, we evaluated scVAG in terms of noise and imbalanced clustering distributions using 8 challenging synthesis datasets. scVAG demonstrated higher performance than baseline methods across adjusted Rand index, cosine similarity, and normalized mutual information metrics, suggesting potential advantages for single-cell clustering tasks. Visualizations of the learned cellular embeddings further demonstrate scVAG's capacity to recover interpretable biological structures. As shown in t-SNE projections, scVAG embeddings clearly distinguish cell subtypes in a manner consistent with ground truth labels, accurately recovering known cell types. In summary, the proposed scVAG architecture establishes a robust framework for elucidating cell taxonomies from noisy, high-dimensional transcriptomic data. ScVAG enables sensitive and accurate clustering at a resolution of single-cell types by combining global data distribution learning with local graph-based feature extraction. This method facilitates biological discovery across applications like development, disease modeling, and monitoring therapeutic responses. Future directions include extending scVAG to integrate multi-omic single cell datasets for further enhanced cell annotation. Rich information beyond gene expression such as chromatin accessibility, DNA methylation, and protein levels will enable more comprehensive cell state characterization. We will also explore predictive modeling scenarios leveraging scVAG's powerful representations, including imputation of missing cell modalities and projection of differentiation trajectories. Limitations of the study For the model to run, three components are required. Initially, there is a preprocessing section, followed by the VAE section, and finally the GATE section. Due to this implementation, the model takes longer to complete than some other techniques, particularly traditional methods like Seurat and SC3. Our experiments show that scVAG requires comparable time and memory to other deep learning-based methods (e.g., scGAC, SCEA), but significantly more than faster traditional approaches. For instance, on the HEK-NIH dataset, scVAG took 4 h to run compared to 0.7 h for Seurat, while using 25 GB of memory. Moreover, if the client does not have a GPU, the VAE and the GATE may take even longer to run. This increased computational demand is a trade-off for the improved clustering accuracy and ability to capture complex cellular relationships demonstrated in our evaluations. Material and methods Data preprocessing In order to achieve robust clustering, we implement a targeted pre-processing approach based on data-driven quality control. The expression matrix d(M′∗N′) consisting of M՛ cells and N՛ genes is converted to d(M∗N), where M