Interpretable modality-aware mapping of gene regulation in single-cell multiomics with scMAGCA Authors: Yunhe Wang, Zeming Zhou, Wei Liu, Peiru Zhang, Yue Cheng, Yanchi Su, Fuzhou Wang, Ka-Chun Wong, Xiangtao Li Published: 2026 ABSTRACT AbstractSingle-cell multi-omics technologies profile multiple molecular layers in individual cells, but existing methods often struggle to integrate transcriptomic, proteomic, and epigenomic measurements into an interpretable representation while preserving relationships among cells. Here, we present the single-cell multi-omics adversarial graph convolutional autoencoder (scMAGCA), which constructs cell graphs and uses adversarial alignment to learn interpretable shared embeddings that capture cellular heterogeneity and regulatory complexity. Across diverse datasets, scMAGCA outperforms existing methods in modality alignment, clustering, and batch correction. In Alzheimer’s disease, scMAGCA resolves neuronal subtypes and regulatory programs that are missed by single-modality analyses. In kidney cancer, it identifies tumor-specific epithelial and endothelial populations and uncovers biomarkers validated by quantitative polymerase chain reaction. These results support scMAGCA as an interpretable framework for resolving complex cell states in disease.Subject terms: Gene expression, Epigenetics analysis, Data mining Introduction Single-omics sequencing has generated vast amounts of individual-modality data, offering valuable insights into the mechanisms of gene regulation and key aspects of cellular heterogeneity1. However, cellular diversity is driven by multiple processes, including proliferation, differentiation, and apoptosis2, and is further influenced by local and distant environmental niches. To fully capture this complexity, it is necessary to analyze multiple molecular layers within the same cell. Recent technological advancements enable the simultaneous profiling of genomes, epigenomes, transcriptomes, and other emerging omics modalities, revolutionizing our understanding of biological mechanisms and genotype-to-phenotype relationships3. Among these, single-cell transcriptomics is the most mature and widely adopted technique; it is often combined with other omics assays to characterize the interplay between gene expression and phenotypic heterogeneity. For instance, the 10x Chromium platform4, coupled with single-cell RNA sequencing (scRNA-seq), uses antibody-derived tags (ADT)5 to quantify cell surface proteins. When integrated with single-cell ATAC-seq (scATAC-seq)6, one can jointly measure chromatin accessibility and gene expression within the same cell. These multi-omics tools have facilitated considerable progress, offering insights into processes ranging from protein dynamics7 to epigenetic variation8. Defining cell identity is pivotal for unveiling biological phenomena, yet it remains challenging due to the intricate molecular and functional heterogeneity of cells. Single-cell multi-omics sequencing addresses this challenge by integrating multiple molecular layers, providing a unified multi-modal perspective on cellular identity. Nonetheless, there is a need to disentangle complex intercellular relationships and accurately classify diverse cell subpopulations. In the single-omics era, many clustering algorithms were developed for scRNA-seq or scATAC-seq separately. Notable graph-based methods (Louvain9, Leiden10) are widely used for their scalability and relative noise tolerance, forming the backbone of pipelines like Seurat11, ArchR12, and Scanpy13. These algorithms typically cluster cells by optimizing modularity on a k-nearest neighbors (KNN) graph after dimensionality reduction, such as PCA14, LSI15. Meanwhile, deep learning methods such as DCA16, scVI17, and scBGEDA18 introduce zero-inflated models, variational autoencoders (VAE), and Kullback-Leibler (KL) divergence to optimize latent representations. These advances yield gains in clustering accuracy in single-omics contexts, but do not inherently address cross-modality dependencies. To overcome these single-omics limitations, multi-omics clustering approaches have emerged to jointly embed different modalities. For example, BREMSC19 uses a Bayesian hybrid model to decouple ADT and RNA in CITE-seq data, while totalVI20 employs a deep VAE for integrative transcriptomic and proteomic representation. SHARE-seq (Single-cell Multiome ATAC + Gene Expression sequencing), focusing on transcriptomic and epigenomic landscapes, introduces further computational challenges due to extreme sparsity and high dimensionality. Methods such as Cobolt21 adopts hierarchical Bayesian VAEs to derive generalized latent embeddings, scMM22 separately models posterior distributions for RNA-seq and ATAC-seq, merging them through a hybrid framework, and GLUE23 employs a graph-guided VAE that incorporates prior regulatory interactions as a graph linking features across modalities. Despite these advances, scMM and Cobolt both use multi-modal VAEs to embed paired gene expression and chromatin data without an explicit graph component, and totalVI is a deep VAE model for CITE-seq that likewise does not exploit any graph structure. VAE-based models like scMM and Cobolt do not include any adversarial component for alignment; they rely only on the VAE’s built-in KL-divergence to a prior, and totalVI optimizes a standard VAE evidence lower bound via stochastic variational inference without graph adversarial training. GLUE does perform adversarial alignment of cell embeddings via a discriminator, but it does so in conjunction with a prior-derived guidance graph. Overall, current multi-omics clustering methods often face critical constraints, as some are designed for specific modality pairs like RNA+ADT or RNA+ATAC, which limits their generality, while others struggle with extreme sparsity or dimensional imbalance across modalities, leading to potential loss of important biological signals. These gaps highlight the need for robust and scalable approaches capable of unifying multiple omics layers, including three or more, in a single, interpretable latent space. In particular, integrating transcriptomic (RNA), proteomic (ADT), and epigenomic (ATAC) modalities within the same cells can yield a more comprehensive understanding of cell identity than any single- or dual-modality approach24,25. Each modality offers unique and complementary insights. Gene expression profiles from scRNA-seq capture a cell’s active transcriptional programs, ADT readouts quantify cell-surface protein markers that define phenotypic traits26, and scATAC-seq maps the chromatin accessibility landscape that underlies gene regulation27. By uniting these three data layers, researchers can uncover complex biological patterns, such as nuanced immune cell states or regulatory interactions, that might be missed by simpler pipelines24,25. However, concurrent analysis of RNA, ADT, and ATAC data presents significant challenges as scRNA-seq typically exhibits substantial dropout and sparsity, ADT measurements are primarily affected by technical noise28, and scATAC-seq profiles are sparse and high-dimensional with fundamentally different signal distributions24. These disparities complicate the direct fusion of tri-modal data, highlighting the need for advanced methods capable of effective tri-modal integration. As multi-omics assays continue to evolve, a variety of graph-based and deep generative strategies now aim to integrate diverse modalities, for instance, Seurat (V4)28, Similarity Network Fusion (SNF)29, and Multi-view Spectral Clustering (MSCLRL)30, build multi-omics similarity graphs for clustering. Seurat (V4) uses a weighted nearest neighbor (WNN) strategy to preserve modality-specific signals, SNF aligns multiple omics matrices to a global similarity structure, and MSCLRL applies spectral clustering on a fused similarity matrix. Deep generative strategies such as scMDC31 and scMVP32 uncover shared and distinct patterns across modalities. However, integrating large-scale, high-dimensional multi-omics data remains a non-trivial task because many traditional clustering tools face scalability issues, while deep neural networks, though powerful for feature extraction, often lack explicit mechanisms to reconcile global cross-modal differences and preserve local cell-cell topology, both of which are critical for accurately resolving complex subpopulations. Bayesian or linear factor approaches, such as MOFA +33 offer stronger interpretability via a linear factor model, but their linear-Gaussian assumptions can be restrictive for count-based, zero-inflated, and over-dispersed single-cell modalities, such as scRNA-seq and scATAC-seq, and they typically do not incorporate an explicit cell-cell graph prior. To address these shortcomings, graph-based methods like DeepMAPS34 have emerged. By centering on message passing over cell-gene graphs, DeepMAPS illustrates the value of explicitly modeling cell-gene relationships for improved modality alignment and preservation of local topology. Here, we introduce scMAGCA, a deep learning framework that integrates adversarial learning, graph convolution, and zero-inflated modeling to jointly represent heterogeneous omics layers. Unlike existing approaches, scMAGCA constructs a multi-modal graph encoder to retain local intercellular relationships in the latent space and employs an adversarial alignment module to unify distributions across different modalities. In addition, a unified zero-inflated negative binomial (ZINB) likelihood explicitly models multi-omics count data, explaining cross-modal variation within a single probabilistic framework. This integrated approach enables scMAGCA to model data sparsity explicitly, learn distribution-invariant embeddings, and preserve fine-grained topological structure in the latent space, thereby extending the modeling capacity of prior frameworks like MOFA+ and DeepMAPS. During pretraining, reconstruction, ZINB, and adversarial objectives harmonize the signals from RNA, ADT, and ATAC; subsequently, a KL-based deep embedding clustering step refines cell-type distinctions in an end-to-end manner. We benchmark scMAGCA against existing computational tools on various RNA+ADT and RNA+ATAC datasets as well as tri-modal TEA-seq data (RNA+ADT+ATAC). Our results show favorable performance in modality alignment, clustering accuracy, and batch-effect removal, while preserving a biologically interpretable latent space that facilitates the discovery of rare or subtle cell populations. Furthermore, scMAGCA reveals cell states and key regulatory relationships, evidenced by its success in detecting disease-specific subpopulations and critical biomarkers in kidney cancer and Alzheimer’s disease. These findings underscore the broad applicability of scMAGCA in the dissecting of complex biological systems and the guide for future disease research and therapeutic strategies. Overview of scMAGCA architecture scMAGCA is a comprehensive, end-to-end deep learning framework designed to harmonize and interpret heterogeneous single-cell omics data. In contrast to single-modality or simpler VAE-based methods, scMAGCA provides a three-pronged integration strategy: (1) graph convolution to preserve cell topology, (2) adversarial learning to align multi-omics distributions, and (3) ZINB to model multi-omics count data, providing coherent and interpretable embeddings even in complex data scenarios. By fusing transcriptomic, proteomic, or epigenomic signals within a unified framework, scMAGCA not only captures fine-grained cellular heterogeneity but also uncovers functionally relevant features, facilitating the discovery of disease-associated subpopulations, biomarker candidates, and regulatory landscapes. Demonstrations across RNA+ADT and RNA+ATAC data further underscore its adaptability to diverse biological contexts, including kidney cancer and Alzheimer’s disease. The framework integrates the following major analytical steps (Fig. 1). The architecture of scMAGCA. aMulti-level regulation from epigenetics to proteomics: Single-cell multi-omics data illuminates regulatory processes spanning chromatin accessibility, transcriptomes, and proteomes, providing an integrated view of cellular states. bInput and data preprocessing: Each modality (RNA, ATAC, ADT) is preprocessed independently; highly variable features are selected, normalized, and concatenated to form a joint cell-by-feature matrix for integration. A K-nearest neighbor (KNN) cell-cell graph is built to preserve intercellular relationships. c, d,gPretraining phase of scMAGCA:c Graph convolutional encoder: A multi-layer graph convolutional encoder aggregates cell nodes and neighborhood information to yield topology-preserving cell embeddings. dAdversarial stage: An adversarial stage with a discriminator aligns the embedding distribution to a prior, reducing modality shifts already during pretraining. gDecoder: Two decoders reconstruct the inputs from the latent space: a linear decoder operating on the preprocessed data and a ZINB decoder for the concatenated raw data matrix after feature selection. e–gTraining phase of scMAGCA:eGraph convolutional encoder: The multi-layer graph convolutional encoder is warm-started from the pretraining weights. fClustering stage: Initialize centroids in latent space; refine with a target distribution and minimize a KL objective to sharpen assignments while preserving structure. gDecoder: Shared with pretraining phase. hApplication and biological validation: scMAGCA accurately integrates and clusters multi-modal data, including tri-modal TEA-seq (RNA+ADT+ATAC) while effectively correcting batch effects. It delineates neuronal-subtype heterogeneity in Alzheimer’s disease and uncovers kidney-cancer biomarkers validated by qPCR, highlighting their roles in renal tumorigenesis and their potential clinical relevance. iFunctional genomics interpretation: Weight matrices in the graph convolutional network (GCN) layers capture key biological patterns. Via tracing backward from the latent space to the input layer, scMAGCA identifies critical biomarkers-genes, chromatin-accessibility peaks, or ADTs-that shape its latent representations. Schematics in (a, b, d, f, and h and i) created in BioRender. Zhou, Z. (2026) https://BioRender.com/t2ctudy. Initially, raw omics data are filtered to retain highly variable genes or peaks, then normalized and log-transformed to mitigate technical noise. These refined features are concatenated into a cell-gene count matrix, forming the foundational input for subsequent modeling. To preserve the cellular topology and exploit local biological relationships, scMAGCA constructs a cell-cell graph from the integrated count matrix using the K-nearest neighbors (KNN) algorithm. Each node represents a cell, and edges capture cell-cell proximity based on multi-omics similarity. A graph convolutional network (GCN) processes this graph, iteratively reducing the dimensionality of high-dimensional data while amplifying omics-specific signals. BatchNorm1d and PReLU layers stabilize training and preserve critical topological features. After that, in the latent space, scMAGCA employs two decoders. A linear decoder aims to reconstruct the count matrix with MSE loss, ensuring representation fidelity. To address zero-inflation and dispersion typical of single-cell data, a ZINB decoder explicitly models dropout events. This ZINB component is especially crucial for scRNA-seq, capturing sparse gene expression patterns and mitigating biases introduced by technical dropouts. Then, scMAGCA incorporates a discriminator at the bottleneck layer, implementing adversarial training to align the aggregated posterior distribution of latent embeddings with a specified prior. Unlike simpler approaches that ignore inter-omics distribution mismatch, this adversarial module refines the latent space so that it is both biologically coherent and robust to modality-specific noise. As a result, scMAGCA enhances cluster separability and maintains intrinsic data structure across diverse single-cell omics. For final clustering, scMAGCA applies a Kullback-Leibler (KL) divergence-based optimization. K-means initialization locates initial cluster centroids in the pre-trained latent representation, and the model jointly updates these centroids alongside reconstruction, ZINB, and clustering losses. This iterative process automatically refines boundaries between cell subpopulations, achieving superior delineation of subtle cellular states. scMAGCA accurately resolves cellular heterogeneity across diverse single-cell multi-omics datasets To evaluate the performance of scMAGCA in accurately identifying cell populations from diverse multi-omics platforms, we conducted a comprehensive benchmarking analysis across 31 RNA+ADT and RNA+ATAC datasets. Specifically, we evaluated scMAGCA against fourteen multi-modal integration methods: Seurat28, MOJITOO35, totalVI20, SCOIT36, scArches37, scVAEIT38, CiteFuse39, MIDAS40, scMM22, DeepMAPS34, scMDC31, MOFA+33, BREMSC19, and scDeepCluster41. We first evaluated scMAGCA using fourteen RNA+ADT datasets, encompassing 713 to 30,672 cells derived from various tissues, including human peripheral blood mononuclear cells (PBMCs), mucosa-associated lymphoid tissue tumors, spleen and lymph nodes. Performance was assessed using Normalized Mutual Information (NMI), Adjusted Rand Index (ARI), Adjusted Mutual Information (AMI), and Accuracy (ACC), employing ground-truth cell type labels. Following our evaluation protocol, five representative RNA+ADT datasets (Dataset IDs 1–5) are used only as a development set to select default hyperparameters using label-free criteria. In Fig. 2a, we report per-dataset AMI, NMI, ARI, and ACC on all fourteen RNA+ADT datasets for transparency, whereas the aggregated benchmark summaries (average rank and overall p-values) are computed exclusively on the remaining nine held-out RNA+ADT benchmark datasets. As shown in Fig. 2a and Supplementary Table 1, across the nine held-out RNA+ADT datasets, scMAGCA attained the top or tied-top average rank on all four metrics, outperforming competing methods. Among the baselines, scMDC and scDeepCluster achieved the closest average ranks to scMAGCA on several metrics, indicating their capability for clustering RNA+ADT data. By contrast, MOFA+, as a probabilistic factorization framework lacking explicit topological priors and nonlinear feature learning, performed suboptimally in our benchmarks. Significance was also evaluated across the four metrics, the resulting p-values from the t test in Fig. 2a and Supplementary Table 2 computed on the nine held-out datasets show that scMAGCA achieves statistically significant improvements over most competing approaches, supporting its superiority in resolving cellular heterogeneity within RNA+ADT datasets. Benchmarking of scMAGCA in terms of cell clustering on multiple real-world scMulti-omics data. a, b Comparative evaluation of clustering performance across 14 labeled RNA+ADT (a) and 14 RNA+ATAC (b) datasets using the average rank on the held-out benchmark datasets, Adjusted Mutual Information (AMI), Normalized Mutual Information (NMI), Adjust Rand Index (ARI) and Accuracy (ACC) evaluation metrics on all the datasets, p-values on the held-out benchmark datasets, and running times (log-scale) on all the datasets for scMAGCA and competing methods. (ns: p > 0.05, *: p < 0.05, **: p < 0.01; ***: p < 0.001) Ten representative datasets were used for hyperparameter tuning (RNA+ADT: IDs 1--5; RNA+ATAC: IDs 19–23), while IDs 6–14 and 24–32 served as the held-out benchmark datasets. A two-sided t test was applied to evaluate the performance of scMAGCA relative to competing methods across held-out benchmark datasets (RNA+ADT, n = 9 datasets; RNA+ATAC, n = 9 datasets); p-values were not adjusted for multiple comparisons. Exact p-values, test statistics, degrees of freedom, effect sizes and confidence intervals in panels (a) and (b) are provided in Supplementary Tables 2 and 4, respectively. c UMAP (Uniform Manifold Approximation and Projection) visualizations of the embeddings obtained by scMAGCA and other competing multi-omics integration methods on the GSE128639 dataset. d Comparative evaluation of clustering performance across three unlabeled RNA+ATAC datasets using evaluation metrics including Average Silhouette Width (ASW), Davies-Bouldin index (DB), and Calinski-Harabasz index (CH), and p-values between scMAGCA and other competing methods. (ns: p > 0.05, *: p < 0.05, **: p < 0.01; ***: p < 0.001) A two-sided t test was used, and p-values were not adjusted for multiple comparisons (n = 3 datasets). Exact p-values, test statistics, degrees of freedom, effect sizes and confidence intervals are provided in Supplementary Table 5. e Ranked eRegulon Specificity Scores (eRSS) across cell types for the pbmc_10X_public dataset, highlighting the top transcription factors (TFs) with the highest specificity in each cluster. Source data are provided as a Source Data file. We next tested scMAGCA on fourteen well-annotated RNA+ATAC datasets, comparing it against nineteen multi-omics integration methods Seurat28, MOJITOO35, SCOIT36, scVAEIT38, DeepMAPS34, scAI42, MOFA+33, scMVP32, Schema43, MIDAS40, multiDGD44, GLUE23, scMM22, scMDC31, Cobolt21, MultiVI45, MIRA46, scCross47, and scDeepCluster41. Similarly, we restricted hyperparameter selection to a development set comprising five RNA+ATAC datasets (Dataset IDs 19–23). Consequently, the headline aggregate summaries in Fig. 2b represent only the nine held-out benchmarks, although we provide the full breakdown of AMI, NMI, ARI, and ACC across all fourteen datasets for completeness. As illustrated in Fig. 2b and Supplementary Table 3, across the nine held-out RNA+ATAC datasets, scMAGCA achieved the best average ranks on ARI and ACC, and ranked second on AMI and NMI, showing consistently solid clustering performance on held-out benchmarks. Relative to those baselines, scDeepCluster and MOJITOO demonstrated competitive performance across AMI, NMI, and ACC, reflecting reliable clustering accuracy. MultiVI also captured cell-cell similarity effectively, yielding competitive clustering outcomes. Several methods exhibited trade-offs: scMDC posted an excellent average ARI but was less competitive on the other metrics, whereas SCOIT prioritized speed—achieving the fastest runtime—at the expense of accuracy. Notably, GLUE’s incorporation of prior biological knowledge yielded strong performance, highlighting that biologically informed coupling of modalities is key to precise cell-type identification. The p-values on the nine held-out RNA+ATAC datasets (Fig. 2b and Supplementary Table 4) demonstrate that scMAGCA achieves statistically significant improvements compared with a range of competing methods. To elucidate the learned representations, we visualized low-dimensional embeddings on the GSE128639 dataset (Fig. 2c), comparing scMAGCA with baseline methods. The GSE128639 dataset, publicly available in the Gene Expression Omnibus (GEO), consists of 30,672 human peripheral blood mononuclear cells profiled using Cellular Indexing of Transcriptomes and Epitopes by sequencing (CITE-seq). These cells are categorized into 27 distinct cell types. Our results in Fig. 2c demonstrate that scMAGCA resolved finer cellular distinctions that were undetectable using individual omics layers alone. Specifically, the UMAP projections produced by scMAGCA reveal finer cellular distinctions. In particular, it achieved clear separation between CD8 Naive and CD4 Memory cells (blue circles), a resolution that could not be achieved using raw ADT or RNA data alone. By contrast, several baselines show distinctive, method-induced patterns. scDeepCluster produces numerous outliers far from the main manifolds, fragmenting same-class cells into multiple small, disconnected clusters; the embedding appears radial with pronounced elongation. BREMSC does not preserve local neighborhood structure well, and phenotypes that should be compact appear overly diffuse. SCOIT and scArches show substantial cluster overlap with mixed labels, indicating limited separability under the same UMAP protocol. Similarly, totalVI fails to resolve certain cell types in specific regions, whereas MIDAS, despite being effective overall in terms of separation, yields indistinct boundaries between adjacent clusters. In addition, CiteFuse maintains continuity between Memory B cells and Naive B cells, suggesting it preserves biological relationships rather than merely projecting the count matrix to a low-dimensional space. scMM also recovers this association but with lower boundary resolution, revealing certain limitations. scMAGCA effectively integrates complementary signals from ADT and RNA data, enabling a more comprehensive capture of the topological relationships between cell populations. This enhanced resolution allows for precise distinction of CD4 Naive, CD8 Naive, CD4 Memory, and NK cells, as well as subtle differences between the CD8 Effector1 and CD8 Effector2 clusters. In particular, scMAGCA successfully differentiates Naive B cells from Memory B cells (red circles), preserving the biological continuum between these populations. Naive B cells serve as precursors in the adaptive immune response, while memory B cells can rapidly mount recall responses upon re-exposure to the same antigen. The extended results of Supplementary Figs. 1–8 further demonstrate scMAGCA’s superior cell-type discrimination in multiple datasets. Following that, to further demonstrate scMAGCA’s applicability in unlabeled single-cell multi-omics data, we applied it to three unlabeled paired scRNA-seq and scATAC-seq datasets, including GSM4949911, pbmc_10X_public, and 10X-Multiome-pbmc10k. These datasets were used to benchmark scMAGCA against nineteen single-omics and multi-omics integration methods. Integration performance was assessed using three widely used evaluation metrics: Average Silhouette Width (ASW), where higher values indicate more compact and well-defined clusters; Davies-Bouldin (DB) index, where lower values indicate better-separated clusters; Calinski-Harabasz (CH) index, where higher values signify a more optimal clustering structure. As shown in Fig. 2d, scMAGCA generally achieved the best performance across all three unlabeled RNA+ATAC datasets, demonstrating its ability to maintain cell-cell consistency without prior annotations. Furthermore, p-values obtained from the t test confirmed that, across those metrics, scMAGCA’s improvements over almost all competing methods were statistically significant (Supplementary Table 5). These results highlight scMAGCA’s superior clustering performance in fully unsupervised settings, reinforcing its applicability to diverse single-cell multi-omics studies where predefined labels are unavailable. Having already benchmarked scMAGCA on the pbmc_10X_public dataset and shown that it achieves the best unsupervised clustering performance (Fig. 2e), we next asked whether those computational gains translate into biologically informative insights. To this end, we revisited these 10,085 cells, now focusing on downstream biological annotation. scMAGCA successfully identified five distinct cell clusters based on cell-type-specific marker genes, corresponding to CD14 Mono (CD14, LYZ, S100A8), CD4 TCM (CD4, CD27, IL7R), CD8 Naive T (CD8A, CD8B, CD45RA), Naive B (CD19, MS4A1, IGHD), and natural killer (NK) (NCAM1). Within each subpopulation, we assessed both differentially expressed genes (RNA) and differentially accessible chromatin peaks (ATAC), then used SCENIC +48 to calculate the Regulon Specificity Score (RSS), which quantifies how selectively a transcription factor (TF) regulates each cluster (RSS values range from 0 to 1). As illustrated in Fig. 2e, the top five TFs for each cluster frequently co-bind shared enhancers, reinforcing the notion that transcriptomic and epigenomic features converge on common regulatory sites. For naive B cells, EBF1, SPIB, and POU2AF1 emerged as key drivers of early B-cell lineage commitment. In CD8 naive T cells, KLF6 and FOXO1 were highlighted, consistent with their established roles in T-cell quiescence and homeostatic survival. CD4 TCM cells showed high specificity for MAF and GATA3, each implicated in T-helper subset differentiation. Meanwhile, NK cells featured MYBL1, EOMES, and TBX21, while KLF12 and TFEC dominated in CD14 Mono cells. In naive B cells, we revealed two prominent open chromatin regions upstream of the BTNL9 (chr5:180990109-180990609 and chr5:180975299-180975799), which exhibit higher accessibility than in other cell types (Supplementary Fig. 9). These regions are putative EBF1-binding sites, suggesting a regulatory role in BTNL9 expression specific to naive B cells. In addition, ATAC-seq analysis also detected strong peaks at these loci. Previous studies have reported that BTNL9, as a potential prognostic biomarker, is associated with the level of naive B cell infiltration in the tumor microenvironment, where its low expression correlates with reduced infiltration49. This association highlights the importance of BTNL9 in the tumor immune microenvironment and its potential impact on patient prognosis. Therefore, further investigation into the regulatory mechanisms of EBF1 and its interaction with BTNL9 could provide critical insights into the functional role of BTNL9 in tumor biology. Beyond clustering accuracy, we assessed computational scalability by measuring running times and memory on 31 RNA+ADT and RNA+ATAC datasets (Fig. 2a, b and Supplementary Figs. 10, 11) among various multi-omics integration methods. In Fig. 2a, scMAGCA achieves a third-place average runtime rank on the 14 RNA+ADT datasets, maintaining high computational efficiency while delivering superior clustering accuracy. On the 14 RNA+ATAC datasets, the runtime of scMAGCA is mid-range, neither among the quickest nor the slowest, yet its clustering accuracy remains competitive. As shown in Supplementary Fig. 10, the extended results confirm this trend for scMAGCA across all unlabeled paired scRNA-seq and scATAC-seq benchmarks against the other thirteen deep-learning integration baselines. Supplementary Fig. 11 shows that scMAGCA’s computational requirements are moderate and on par with other multi-omics integration methods. For example, on a typical RNA+ADT dataset of 5.5k cells (10X5kpbmcTotalSeq), scMAGCA completes integration in a few minutes and requires under 2 GB of memory, comparable to methods like totalVI. On an RNA+ATAC dataset of 5.9 k cells (GSE156478_Stim), scMAGCA completes integration using less than 4 GB of memory. Moreover, to showcase the cross-modality versatility of scMAGCA, we extended our analysis to paired single-cell RNA-seq and DNA methylation data using two representative datasets: the Cheow dataset (177 human cells during reprogramming) and GSE140493 (3,898 single nuclei from human frontal cortex). Full details on the evaluation are provided in the Supplementary Note 2. Together, these results highlight the broad compatibility of scMAGCA across different data platforms (CITE-seq, SHARE-seq, REAP-seq, DOGMA-seq), sample sizes (from hundreds to tens of thousands of cells), and species (human and mouse). By integrating RNA+ADT and RNA+ATAC data, scMAGCA provides a unified framework capable of harmonizing diverse single-cell datasets, a key advantage over approaches that focus on a single model or lack cross-platform validation. scMAGCA ensures robust cross-batch integration while preserving biologically interpretable single-cell representations Cross-batch heterogeneity often obscures true biological signals in single-cell analyses, as technical discrepancies from multiple experimental runs can introduce artificial variation. Achieving robust batch correction without sacrificing cell-type fidelity is thus critical for reliable downstream interpretations. To determine whether scMAGCA achieves this balance, we tested its performance on five multi-batch multi-omics datasets (GSE164378, GSE175702, GSE163120, GSE156478 (RNA+ADT), and GSE156478 (RNA+ATAC)), evaluating its capacity to harmonize data from different batches while maintaining biologically meaningful single-cell representations. We compared scMAGCA against eight multi-omics integration algorithms totalVI20, scArches37, scVAEIT38, MOFA+33, MIDAS40, DeepMAPS34, Seurat28, and scMDC31 were applied to RNA+ADT datasets. While thirteen scMulti-omics integration algorithms: scVAEIT38, MOFA+33, MultiVI45, scMoMaT50, MIRA46, MIDAS40, multiDGD44, GLUE23, DeepMAPS34, Seurat28, scMDC31, Cobolt21, and scCross47 were applied to the RNA+ATAC dataset. To quantitatively assess cross-batch integration, we employed three categories of evaluation metrics: (1) Biological conservation metrics: Isolated labels, KMeans NMI, KMeans ARI, Silhouette label, and cLISI; (2) Batch correction metrics: Silhouette batch, iLISI, KBET, and GC; (3) Aggregate scores: the unweighted mean of batch correction and biological conservation metrics, together with a total score computed as a 40:60 weighted mean (batch correction: biological conservation)51. As shown in Fig. 3a and Supplementary Table 6, scMAGCA outperforms competing methods in terms of batch correction and biological preservation metrics on GSE164378, which contains 161,764 cells from two batches and 31 cell types. Notably, scMAGCA excels in the aggregate score of batch correction and biological conservation, indicating its ability to effectively harmonize data from different batches without compromising biological variability. Importantly, while some competing methods eliminate batch effects at the expense of biological variation, scMAGCA achieves optimal batch integration while maintaining the integrity of cell type distinctions. To further investigate batch integration, we visualized the low-dimensional embeddings generated by scMAGCA and competing methods (Fig. 3b). As observed from Fig. 3b, it is evident that scMAGCA and competing methods successfully identified all major cell clusters, while scMDC failed to produce clustering results consistent with known cell populations, highlighting its limitations in multi-batch data integration. Among competing methods, DeepMAPS achieved good clustering performance, with batch-correction results broadly in line with other methods. In addition, MIDAS and totalVI exhibited obvious batch-dependent clustering, separating cells from different batches rather than based on true biological differences. In contrast, scMAGCA successfully integrated cells from both batches into cohesive clusters, demonstrating its superior ability to harmonize multi-modal data while preserving cellular identity relationships. Evaluation of scMAGCA in cross-batch integration and biological interpretability on the CITE-seq data GSE164378. a Overall evaluation of scMAGCA compared to other competing methods. The evaluation is divided into three categories: biological conservation, batch correction, and aggregate score. These metrics assess the ability of each method to maintain biological integrity while correcting for batch effects. b UMAP visualizations of cell embeddings for scMAGCA and other competing methods. Cells are colored by predicted labels (left) and batch labels (right). The UMAP plot demonstrates how well each method preserves biological clustering (based on predicted labels) and corrects for batch effects (based on batch labels), visually comparing the batch integration quality across methods. c Top-3 factor specificity per cell type for scMAGCA and competing methods. The specificity score is computed as one-vs-rest enrichment. Note that a factor may show high specificity in multiple cell types when it is strongly enriched in those types relative to others. Cell-type labels were derived from the clustering results and subsequent marker-gene-based mapping to the GSE164378 reference, and all methods were evaluated on the same set of reference cell types. The number of latent factors equals the dimensionality of each method’s learned embedding, following default settings without manual tuning. The vertical axis shows the mean factor weight in that specific cell type (CD4 TCM, CD4 CTL, CD4 Proliferating, CD8 TCM, CD8 TEM, B naive, CD14 Mono), and the horizontal axis shows the mean factor weight in all other cell types. The comparative performance on the remaining four multi-batch datasets is summarized in Supplementary Fig. 13 and Supplementary Table 6. We can observe that scMAGCA secures the highest total scores across these datasets and attains the highest Graph Connectivity on GSE175702, with similarly strong performance on the remaining RNA+ADT datasets, providing evidence of superior maintenance of cellular topological structure. Then, to assess the biological interpretability of scMAGCA’s clustering results on GSE164378, we evaluated the specificity scores of latent factors associated with individual cell types via one-vs-rest enrichment, following the approach in ref. 52. Fig. 3c and Supplementary Fig. 14 show the top-3 factor specificity of each method across multiple immune cell clusters, including CD4 TCM, CD4 CTL, CD4 Proliferating, CD8 TCM, CD8 TEM, B naive, and CD14 Mono cells. A high specificity score indicates that a model successfully assigns distinct latent factors to individual cell types, reinforcing the biological interpretability of the clustering process. However, a factor can appear highly specific in multiple cell types if it is strongly enriched in each relative to the others. Our analysis shows that scMAGCA attained the highest specificity scores across diverse cell types. Notably, a closer examination of subpopulations revealed that Factor 7 is primarily enriched in CD4 CTL cells, followed by CD8 TEM cells, and then CD4 TCM cells. Despite not being exclusive to CD4-lineage cells, Factor 7 appears to capture a cytotoxic-memory transcriptional program that is shared across the CD4 CTL, CD8 TEM, and CD4 TCM subsets. CD4 CTLs are known to share many gene-expression features with CD8+ cytotoxic/effector T cells53. In fact, previous transcriptomic analyses have highlighted an extensive commonality of cytotoxic and effector-memory gene expression signatures among CD4 CTLs, CD8 CTLs, and effector-memory T cells53. It is therefore biologically plausible-and even expected-that a latent factor like Factor 7 would span multiple T cell subsets that participate in related immune functions. Consistent with this view, our KEGG pathway enrichment analysis of the genes highly loaded on Factor 7 across the CD4 TCM, CD4 CTL, CD8 TEM cell subsets in Supplementary Fig. 15 shows significant enrichment for pathways central to T cell activation and effector function, including T cell receptor signaling pathway, Cell adhesion molecules, and Antigen processing and presentation. These pathways are hallmarks of activated and memory T cells in both the CD4 and CD8 lineages, supporting the notion that Factor 7 reflects a shared effector/memory T-cell program across those subsets. This finding shows that scMAGCA enhances interpretability by capturing shared transcriptional programs across related cell states, uncovering meaningful latent factors that span immune subtypes and revealing functional links between cell types (cytotoxic and memory T cells). Building on these results, to substantiate the complementary strengths of scMAGCA, especially compared with DeepMAPS and MOFA +, we conducted targeted experiments addressing three key challenges in single-cell multi-modal integration: preserving topological structure, aligning heterogeneous distributions across batches/modalities, and yielding interpretable latent factors (see Supplementary Note 4 for details). To fully annotate an example latent factor and demonstrate its biological relevance, we focus on Factor 3 in the CD4 TCM subset as a representative example in Supplementary Fig. 16, illustrating how scMAGCA’s latent factors can be interpreted in a biologically meaningful way. As shown in Supplementary Fig. 16a, Factor 3 is highly specific to the CD4 TCM cluster. Concretely, in the CD4 TCM subset, Factor 3 can be summarized as “Factor 3 → top genes (MAL, IL7R, BCL2, ITGB1, JUNB) → GO terms related to CD4-positive alpha-beta T cell and T-helper differentiation (FDR < 0.05)”. We first identified the top 100 gene contributors (Supplementary Table 7) to Factor 3 in CD4 TCM cells, including MAL, IL7R, BCL2, ITGB1, and JUNB. These genes are closely associated with central memory CD4 T cell biology. GO enrichment analysis of this gene set in Supplementary Fig. 16b–d showed that Factor 3 is dominated by GO terms describing CD4-positive alpha-beta T cell differentiation involved in immune response and T-helper cell differentiation, both significantly enriched at a false discovery rate (FDR) below 0.05. These processes indicate that Factor 3 captures the core differentiation programs that generate and maintain CD4 T helper cells with central memory features, consistent with the survival, trafficking, and activation signatures observed in CD4 TCM cells54. In parallel, the cytosolic ribosome and cytosolic large ribosomal subunit were enriched (FDR < 0.001), consistent with a biosynthetically competent state that enables rapid translation of effector molecules upon recall stimulation in CD4 TCM cells. Molecular function terms, including protein tyrosine kinase binding, tumor necrosis factor receptor binding, and chemokine binding (FDR < 0.05), further highlight signaling modules that integrate activation, survival, and migratory cues. Taken together, these GO terms support the view that Factor 3 reflects a CD4 TCM program that preserves homeostasis and tissue trafficking while maintaining a heightened capacity for protein synthesis and signal transduction, thereby preparing central memory T cells for rapid effector responses upon restimulation. We next assess computational scalability-runtime and memory by modality (Supplementary Fig. 17). Overall, on four RNA+ADT datasets (Dataset IDs: 15–18), methods fall close together on the log-runtime axis and generally use only a few gigabytes of memory, scMAGCA is mid-range in log-runtime and uses a few GB, while several methods are faster, Seurat stands out on the largest dataset (GSE164378) with much higher memory use. On the RNA+ATAC dataset (Dataset ID: 33), log-times are similarly compact, seconds-level across the board, while memory spreads more. Several frameworks remain under 1 GB, a few are substantially heavier; scMAGCA falls toward the higher side of memory in this modality, yet maintains seconds-level log-runtime. Furthermore, to address missing modalities in multi-omics integration, we apply and further adapt scMAGCA to distinct missingness scenarios, as detailed in Supplementary Note 5, including scenarios where certain cells lack specific omics layers (Supplementary Fig. 19), multi-omics mosaic integration (Supplementary Figs. 20, 21), and preliminary cross-modal prediction (Supplementary Fig. 22). scMAGCA enables robust tri-modal data integration Although dual omics approaches (RNA+ADT or RNA+ATAC) have substantially advanced single-cell analyses, fully integrating transcriptomic, proteomic, and epigenomic signals can uncover deeper immunological insights often missed by simpler pipelines. Nevertheless, analyzing three distinct modalities in parallel poses significant challenges. RNA exhibits high dropout and overdispersion; ADT has lower dropout than RNA, but substantial ambient background and heavy-tailed noise; and ATAC is sparse and follows a fundamentally different signal distribution. Existing methods often struggle to harmonize these modalities, leading to partial feature alignment or the loss of key biological signals. In contrast, scMAGCA leverages a graph-convolutional backbone with adversarial alignment, unifying the modalities into a coherent latent space and explicitly modeling dropout using a ZINB-based decoder. To illustrate these advantages, we applied scMAGCA to a human PBMC TEA-seq dataset (GSE158013), profiling 7084 cells across transcriptomic, proteomic, and epigenomic dimensions. We first compared scMAGCA with Seurat in six input configurations: RNA only, ADT only, ATAC only, RNA+ADT, RNA+ATAC, and RNA+ADT+ATAC, verifying that each added modality improved clustering performance (Fig. 4a). Moreover, scMAGCA consistently outperformed Seurat integrations, indicating that an adversarial graph-based approach captures multi-modal variance more effectively. Tri-modal integration and biological insight evaluation of scMAGCA on the human PBMC TEA-seq dataset GSE158013. a Comparison of clustering performance between scMAGCA and Seurat across different omics modalities on the TEA-seq dataset GSE158013. b UMAP overlays (top) show RNA (NCAM1, CD8A, ICOS) and corresponding ADT proteins (CD56, CD8a, CD278). Violin plots (bottom) show within-cell-type Spearman correlations between normalized RNA and ADT for matched marker pairs (blue) versus mismatched negative-control pairs (gray). Distributions were generated via 100 bootstrap resampling iterations of individual cells, demonstrating that concordance is specific to biologically related pairs. Box plots within violins indicate the median (center line), interquartile range (IQR; box limits), and whiskers extending to a maximum of 1.5 × IQR from the hinge, with outliers plotted individually. The exact cell numbers (n) used for resampling each cell type are: B cell (1363), CD8 T cell (668), CD8 + MAIT (276), Naive CD4 T cell (2437), Natural killer cell (511), Neutrophil lineage (757), and T follicular helper cell (1072). c Joint UMAP colored by the ATAC accessibility fraction. We define the ATAC accessibility fraction as the fraction of peaks detected in each cell relative to the total number of detected peaks across all cells. This visualization highlights heterogeneity in chromatin accessibility levels within the joint embedding. d Zoom-in of neutrophil lineage and B cell populations colored by the ATAC accessibility fraction, revealing two subgroups with distinct accessibility levels: high chromatin accessibility (HCA) and low chromatin accessibility (LCA). The separation of HCA and LCA subclusters indicates that incorporating chromatin accessibility via scMAGCA uncovers finer-grained regulatory heterogeneity and enables further stratification within these populations. e Top contributing genes for Factor 6 in HCA B cells. The vertical axis shows the mean factor weight in that specific cell type (HCA B cell), and the horizontal axis shows the mean factor weight in all other cell types. f GO and KEGG enrichment analysis of the top 100 genes from the HCA B cell-specific latent factor 6 identified by scMAGCA using the `enrichGO' and `enrichKEGG' functions. FDR, Benjamini-Hochberg adjusted P values, were transformed as −log10(FDR) for visualization. By annotating clusters using the ‘FindAllMarkers’ function, we identified diverse immune subsets-B cells (FCRL1), naive CD4 T cells (LEF1), T follicular helper cells (ICOS, ADT: CD278), CD8 T cells (CD8A, ADT: CD8a), natural Killer (NK) cells (NCAM1, ADT: CD56), CD8+ MAIT cells (CD8A, SLC4A10), and neutrophil lineage cells (CSF3R), which formed distinct UMAP clusters (marked by dashed box in Fig. 4a). Analyzing RNA+ADT data revealed both concordance and divergence between transcript levels and their corresponding protein products. Specifically, the marker genes NCAM1 (NK cells), CD8A (CD8 T cells), and ICOS (T follicular helper cells) are specifically expressed in their respective clusters. We then compute the RNA-ADT Spearman correlation within each annotated cell type and report the full correlation distributions (Fig. 4b). The distributions are obtained from within-cell-type bootstrap resampling over cells with replacement, using 100 resampling iterations, yielding a distribution of Spearman correlation coefficients rather than a single pooled estimate across all cells. We also include mismatched gene/protein pairs as negative controls (CD8A-CD80, ICOS-HLA-DR, NCAM1-CD319), which are centered near zero across cell types, establishing the baseline noise level. Importantly, the results show that correlations are not uniformly high across cell types. Lineage and activation markers exhibit elevated correlations primarily in biologically relevant populations, such as CD8A-CD8a in CD8 T and CD8+ MAIT cells and NCAM1-CD56 in NK cells. We note that the absolute maximum correlation does not necessarily coincide with the cell type exhibiting the highest mean expression, because correlation quantifies co-variation rather than co-enrichment: when a marker is uniformly high within a lineage, leading to a limited dynamic range, or when it is affected by modality-specific noise such as RNA dropout or ADT background signal, within-type correlations can be attenuated even though the marker remains highly specific at the expression level. We additionally include negative-control gene-protein pairs (CD8A-CD80, ICOS-HLA-DR, NCAM1-CD319) in the same plot (gray violins). These controls remain centered near zero across cell types, providing an empirical noise floor and indicating that the positive concordance observed for matched marker pairs is not a trivial consequence of global scaling, shared normalization, or model regularization, but instead reflects pair-specific biological correspondence. These analyses indicate that RNA and ADT provide biologically coherent signals whose concordance is cell-type specific, supporting the rationale for integrating both modalities rather than relying on a single view alone. However, RNA+ADT integration alone still left certain subpopulations unresolved. Incorporating the ATAC dimension via scMAGCA uncovered additional regulatory nuances, particularly within the neutrophil lineage and B cells (Fig. 4c). We define the ATAC accessibility fraction as the fraction of peaks detected in each cell relative to the total number of detected peaks across all cells. In addition, we provide RNA- and ADT-based feature ratios on the joint embedding in Supplementary Fig. 23. While these RNA and ADT-based metrics were comparable across these cells and did not resolve finer subclusters within the neutrophil lineage or B cells (Supplementary Fig. 23), the ATAC accessibility fraction revealed pronounced differences. Specifically, we observed a clear separation between high-accessibility (HCA) and low-accessibility (LCA) subgroups (Fig. 4d), indicating that incorporating the ATAC modality via scMAGCA uncovers additional regulatory heterogeneity and enables finer stratification of these populations. Moreover, to quantitatively assess whether strong alignment is achieved at the expense of label-defined structure and to quantify per-modality variance retention in the joint embedding, we report scaled cLISIand cASW, cross-modal neighborhood correlation scores (cell-wise similarity and global structure score35), and per-modality variance retention in Supplementary Note 6. We then analyzed the scMAGCA-derived latent factors within the B cell compartment (Fig. 4e and Supplementary Fig. 25a). This analysis showed that Factors 14 and 17 represent pan-B cell programs shared by both subsets, whereas Factor 6 is specific to HCA B cells and Factor 5 is specific to LCA B cells. We explicitly link enriched GO terms and KEGG pathways for the top 100 genes associated with the HCA and LCA B cell-specific latent factors (Supplementary Tables 10, 11). These associations are illustrated in Fig. 4e, f and Supplementary Fig. 25 for HCA B cells and LCA B cells, respectively. As shown in Fig. 4f, for HCA B cells, GO and KEGG enrichment of Factor 6 top-loading genes shows significant over-representation of GO terms related to MHC class II-associated protein complexes and immune receptor functions, such as MHC class II protein complex binding and immune receptor activity, as well as the KEGG pathways Antigen processing and presentation and Hematopoietic cell lineage. These enrichments indicate that Factor 6 delineates an antigen-presenting, immune-activated B cell program in HCA B cells, characterized by prominent MHC II machinery and immune-receptor signaling. Consistent with the expression of B cell marker genes, this subset likely represents an immunologically active antigen-presenting B cell population, potentially with immunoregulatory features, whereas any enhancement of differentiation potential would require further evaluation using cell-cycle- and germinal-center-related pathways. In LCA B cells, Factor 5 is characterized by GO and KEGG enrichment of its top-loading genes, as shown in Supplementary Fig. 25b. Within the BP category, Factor 5 is enriched for processes related to small GTPase-mediated signal transduction and MHC class II-dependent antigen processing and presentation. CC enrichment highlights the MHC class II protein complex and related Golgi-associated compartments, while MF enrichment emphasizes immune receptor activity and MHC class II receptor activity. At the pathway level, Factor 5 is associated with KEGG categories such as Intestinal immune network for IgA production and Allograft rejection. Collectively, these results indicate that this LCA B cell subset is strongly enriched for MHC class II-mediated antigen-processing and immune-receptor-related functions, consistent with an antigen-presenting, immune-activated B cell program that is implicated in inflammatory and autoimmune immune responses. Together, these analyses directly connect the enriched GO and KEGG terms to distinct latent dimensions derived from scMAGCA and to the HCA B and LCA B cell type-specific signatures. Collectively, these findings demonstrate that all three modalities (RNA for transcriptional programs, ADT for protein-level confirmation, and ATAC for chromatin accessibility) jointly contribute to a deeper understanding of immune cell heterogeneity. scMAGCA integrates these distinct signals within a graph-based adversarial framework, enabling the detection of subtle regulatory signatures and functional subpopulations, while avoiding the partial or conflicting inferences that can arise from dual-modality or single-modality approaches alone. Such a multi-layered perspective is particularly advantageous in dissecting PBMC compositions, uncovering epigenetically defined sub-clusters, and advancing our knowledge of immune regulation and potential disease mechanisms. Hyperparameter tuning and ablation study of scMAGCA To obtain a label-free global hyperparameter configuration for scMAGCA that generalizes across diverse single-cell multi-omics datasets, while avoiding circular evaluation and excessive computational cost, we first defined ten representative development datasets: five RNA+ADT datasets (Dataset IDs: 1–5) and five RNA+ATAC datasets (Dataset IDs: 19–23). All tuning and sensitivity analyses below were performed without using any cell-type labels, and with all non-target settings kept fixed to isolate the effect of the tested choice. For each candidate configuration, we evaluated the learned representations using only unsupervised clustering criteria: ASW, DB, and CH. Because these metrics have different scales and optimization directions, we summarize performance via an average-rank aggregation. Concretely, for each dataset and each metric, we rank candidate settings (rank 1 is best; higher ASW, CH and lower DB are preferred) and then average the ranks across all development datasets and the three metrics to obtain a single scale-free score for each candidate. Using the development protocol and label-free evaluation scheme, we extended our sensitivity analysis to the KNN neighborhood size and bottleneck dimension. We evaluated neighbor values k ∈ {10, 20, 30, 40, 50} and latent dimensions dim∈{8,10,16,32,64}, ultimately selecting k = 30 and dim=32 as the optimal settings (Supplementary Fig. 26 (a-b)). In addition, we investigated architecture sensitivity across 4 × 3 = 12 encoder–decoder configurations, formed by combining four encoder architectures ({m, 32}, {m, 64, 32}, {m, 256, 64, 32}, and {m, 1024, 256, 64, 32}) with three decoder architectures ({32, m}, {32, 512, m}, and {32, 512, 1024, m}). Based on this comparison, we selected the combination of encoder {m, 1024, 256, 64, 32} and decoder {32, 512, 1024, m} with the lowest average rank as the global default (Supplementary Fig. 26c and Supplementary Table 12). On the development set, we further performed a comprehensive grid search over 27 candidate combinations of the loss weights α, β, and γ (Cases1–Cases27 in Supplementary Table 19), which control the contributions of the reconstruction loss, ZINB loss, and adversarial loss, respectively. Specifically, we swept α ∈ {0.2, 0.5, 1.0}, β ∈ {0.5, 0.8, 1.0}, and γ ∈ {0.001, 0.01, 0.1}. As shown in the scMAGCA panels of Supplementary Fig. 27, Case1 achieves the best overall trade-off across ASW, DB, and CH with the lowest average rank, and is therefore selected as the default (α, β, γ) = (0.2, 0.8, 0.01). After selecting these global defaults on the development set, we fixed them for all remaining datasets and performed no further dataset-specific hyperparameter tuning. To prevent upward selection bias, the ten development datasets are excluded from all headline aggregate benchmark summaries, which are computed only on the held-out datasets. We also conducted additional experiments to evaluate the impact of the key modules in the scMAGCA framework, including the graph encoder, the adversarial alignment module, and the ZINB-based distribution modeling. The ablation study results in Supplementary Note 7 confirm that each component of scMAGCA is essential for its strong performance in integration and clustering. Removing the graph encoder produced the largest performance drop, highlighting the importance of capturing cell-cell relationships. Removing the adversarial alignment module also led to worse alignment between modalities, demonstrating that this module is pivotal for learning unified latent representations. Finally, using a ZINB likelihood for the combined feature space was found to be the most effective way to model multi-omics data; alternative decoding distributions yielded inferior results. Collectively, these findings validate our architectural choices, including the graph-based encoder, adversarial alignment, and unified ZINB decoder, which work in concert to substantially improve integration and clustering performance. scMAGCA can decipher multi-omics regulators of immune function and cell-cell communication Immune cell populations are inherently complex, with gene expression and surface protein markers intricately coordinating key functional processes and intercellular signaling events. Single-modality approaches often fail to capture these nuanced interactions in heterogeneous immune environments. By integrating transcriptomic and proteomic data, scMAGCA provides a holistic view of regulatory networks, enabling precision in pinpointing critical regulators of cell-cell communication. In this context, we applied scMAGCA to dissect the multi-omics underpinnings of immune function, focusing on how the method resolves subtle cell states and unveils immune-modulatory pathways at the single-cell level. Therefore, to evaluate the functional interpretability of scMAGCA, we applied it to the 10Xmalt CITE-seq dataset, which comprises 8412 single cells from diverse immune populations. We first identified 11 functionally distinct cell types using scMAGCA’s multi-omics clustering framework (Fig. 5a). Then, we propose a backward weight attribution framework that progressively selects informative dimensions from the final GCN layer back to the input layer (detailed in Method). Specifically, the importance of each GCN layer’s weights is measured using column-wise standard deviation. By retaining top-weighted dimensions at each step, we trace the propagation path across layers and identify the top 200 input features containing 195 key genes and 5 key ADTs that most influence the latent representation Z, thereby facilitating comprehensive cell-type annotation. This method accounts for the presence of non-linear activations and multi-hop propagation in GCNs, offering a practical and biologically meaningful strategy for interpreting feature contributions under a deep graph architecture. Differential expression analysis (‘FindAllMarkers’ in Seurat) further validated cluster-specific markers (Supplementary Fig. 31a, b show the heatmap of marker ADTs and genes, respectively). Functional genomics interpretability through multi-omics embedded representation of scMAGCA on the CITE-seq dataset 10Xmalt. a UMAP visualization of scMAGCA clustering results on the CITE-seq dataset 10Xmalt. The cell clusters are annotated based on the expression of markers and key genes as well as ADTs, highlighting distinct cell populations within the dataset. b The expression distribution of key genes across eleven cell clusters identified by scMAGCA offers insights into the underlying molecular heterogeneity within the dataset. c The expression distribution of ADTs identified by scMAGCA across the eleven cell clusters. The analysis shows how protein expression profiles correlate with the clustering results, helping to further delineate cellular identities and functions. d Track plots representing the key genes encoding the specific key ADTs of each cell cluster, thereby revealing the molecular heterogeneity across different cell populations. e The KEGG pathway enrichment analysis result reveals significant biological pathways enriched in different cell clusters. This analysis underscores the functional relevance of the cell populations, offering insights into their roles in various biological processes. KEGG pathway enrichment was assessed using a one-sided hypergeometric test, and p-values were adjusted for multiple comparisons using the Benjamini-Hochberg false discovery rate method. f The GSVA result shows the enrichment of specific pathways in each cluster. This quantitatively measures the pathway activity associated with the cell types identified. g The GSEA result shows the enrichment of the T cell receptor signaling pathway and the cell adhesion molecules pathways. These findings align with the GSVA result, supporting the functional significance of these pathways in the dataset. h Cell interaction strength network illustrates the communication between CD8 + T cells and other cell types identified by scMAGCA. The network visualizes how cell-to-cell interactions and signaling pathways mediate cellular communication, especially focusing on CD8 + T cells. i Outgoing and incoming cell communication patterns between different cell groups and signaling pathways. The interactions are depicted as bidirectional links, showing both the source and target of signaling, along with the pathways involved in these communication events. Following clustering with scMAGCA, we annotated 11 distinct immune cell populations by integrating marker genes and key features (Fig. 5a and Supplementary Table 13). scMAGCA’s multi-omics integration enabled the identification of cell populations with fine-grained functional distinctions, such as memory CD4 + T cells and metabolically active B cells, which may be difficult to resolve using single-modality approaches. Notably, eight of the identified key genes overlapped with known marker genes, confirming the robustness of scMAGCA in capturing transcriptional programs relevant to immune function (Fig. 5b). In addition, all key ADTs identified by our selection strategy were independently validated as cell type-specific markers (Fig. 5c), demonstrating that scMAGCA effectively incorporates transcriptomic and surface protein information to refine cell identity resolution in multi-omics single-cell data. Moreover, to assess how scMAGCA translates embeddings into mechanistic insights, we scrutinized the expression correspondence of key genes and their encoded ADTs (Fig. 5d). Notably, IL2RA, TIGIT, and PDCD1 showed direct transcriptional control over their respective protein products CD2555, TIGIT56, and PD-157, all of which were independently identified as key ADTs by scMAGCA. CD8A, a hallmark gene of CD8+ T cells, exhibited strong co-expression with CD8a, demonstrating scMAGCA’s ability to accurately capture gene-protein correspondence in immune cell identity. This gene-protein corroboration validates the framework’s capacity to link RNA-level signals with protein-level identity, while also uncovering gene-protein correlations that suggest potential post-transcriptional regulatory dependencies. To extend beyond single-gene observations and explore broader functional networks, we performed KEGG pathway enrichment analysis (Fig. 5e) and identified the “PD-L1 expression and PD-1 checkpoint pathway in cancer” as a key regulatory axis in the dataset. PDCD1, which encodes PD-1, plays a central role in immune suppression58, where tumor-expressed PD-L1 binds to PD-1 receptors in T cells, inhibiting activation and cytotoxicity. Notably, scMAGCA identified both PDCD1 and PD-1 protein as key characteristics, highlighting its utility in detecting immunosuppressive signaling within the tumor microenvironment. After that, we next performed GSVA to explore the pathways associated with CD8A, a hallmark of CD8+ T cells (Fig. 5f). This analysis identified two key pathways, including T cell receptor (TCR) signaling pathway and Cell adhesion molecules, both critical for immune function. Within the TCR signaling pathway, scMAGCA identified CD8A as a key regulator, improving antigen recognition by binding to MHC I molecules in tumor cells, thus strengthening immune surveillance59. Furthermore, scMAGCA identified PDCD1, which is enriched in this pathway and encodes PD-1, a negative regulator of TCR activation via ITIM and ITSM motifs. This interaction modulates immune responses by preventing excessive T cell activation, which is consistent with its established role in immune checkpoint inhibition60. In parallel, scMAGCA discovered CD8A and TIGIT as pivotal components of the Cell adhesion molecules pathway, mediating T cell-tumor interactions. CD8A, as a co-receptor, facilitates CD8+ T cell binding to tumor-associated antigens (TAAs), reinforcing immune engagement. In contrast, TIGIT contributes to maintaining cell-cell adhesion and tissue integrity56. However, tumors frequently downregulate adhesion molecules such as claudins and occludins, a process involved in cancer metastasis and immune evasion61. Identifying these regulatory interactions underscores the ability of scMAGCA to link transcriptional programs with immune-modulatory functions. To further validate these findings, we performed GSEA (Fig. 5g), which confirmed the enrichment of TCR signaling and cell adhesion pathways, corroborating our GSVA results. The elevated expression of key genes, particularly CD8A, PDCD1, and TIGIT, reinforces their role in modulating T cell responses within the tumor microenvironment. Together, these findings illustrate the ability of scMAGCA to integrate multi-omics data to resolve functional genomic relationships, providing mechanistic insights into immune regulation and tumor cell interactions. Then, we performed cell-cell communication analysis between CD8+ T cells and other immune populations using CellChat (Fig. 5h). As observed from Fig. 5h, we found strong communication signals between CD8+ T cells and Plasma cells, as well as Dendritic cells, suggesting functional cross-talk in immune regulation. Notably, rare cell populations, such as Dendritic cells, were successfully identified by scMAGCA, highlighting its ability to integrate multi-omics data to resolve cell-type interactions beyond what is achievable using single-modality approaches (Supplementary Fig. 32). Next, we examined the molecular pathways driving these interactions and found that the Annexin signaling pathway played a pivotal role in cell-cell communication (Fig. 5i). scMAGCA identified ANXA1 as a key regulator within this pathway, encoding Annexin A1, a protein involved in modulating immune responses. Annexin A1 facilitates anti-inflammatory signaling, regulates T cell function, and mediates dendritic cell migration. Through calcium-dependent phospholipid binding, it modulates immune activation by dampening pro-inflammatory mediators, thereby maintaining immune homeostasis62. Further interaction analysis (Supplementary Fig. 33) revealed that Annexin A1-FPR interactions orchestrate cross-talk between CD8+ T cells and Dendritic cells, a process known to enhance antigen presentation and immune activation. These findings align with previous reports showing that FPR-Annexin A1 signaling facilitates dendritic cell migration and antigen presentation, thereby promoting CD8 + T cell activation during immune responses63. Such an integrated approach combining data-driven clustering, functional pathway analysis, and ligand-receptor inference enables scMAGCA to translate multi-omics embeddings into concrete biological insights. For instance, by linking transcriptional shifts (PDCD1, CD8A) with surface protein markers (PD-1, CD8a) and signaling pathways (TCR, Annexin A1-FPR), scMAGCA illuminates the mechanistic underpinnings of immune modulation at the single-cell level, offering actionable leads for immunological research and therapeutic development. scMAGCA dissects neuronal subtype heterogeneity and uncovers regulatory landscapes in Alzheimer’s disease through multi-omics integration Alzheimer’s disease (AD) is characterized by pronounced neuronal heterogeneity, complicating efforts to unravel the specific cell subpopulations and regulatory mechanisms underlying neurodegeneration. Although single-cell RNA-seq has advanced our understanding of neuronal diversity, transcriptomic analysis alone cannot fully capture the epigenetic regulation that drives disease-specific gene expression changes. To bridge this gap, integrating scATAC-seq with scRNA-seq data is essential, yet current multi-omics approaches often struggle to accurately align these modalities due to high data sparsity and cross-modality heterogeneity. Leveraging its integrative framework that couples adversarial learning and graph convolutional architectures, scMAGCA is well suited to simultaneously resolve fine-grained neuronal subtypes and their epigenetic regulators. We applied it to the RNA+ATAC dataset Multiome_Mouse_Brain_Alzheimers, derived from coronal brain sections of an Alzheimer’s disease (AD) mouse model. This dataset comprises matched scATAC-seq and scRNA-seq profiles from 33,459 single cells, allowing us to investigate transcriptional and epigenetic regulation at the single-cell level. scMAGCA identified 15 distinct cellular clusters using UMAP visualization (Fig. 6a). These annotations were validated based on established cell-type-specific marker genes (Fig. 6b, Supplementary Table 14, Supplementary Fig. 34). Among these populations, excitatory neurons represented the largest fraction, accounting for 25.77% of all cells (Supplementary Fig. 35), emphasizing their critical role in neuronal communication through glutamate-mediated synaptic transmission. Given the well-documented excitotoxicity hypothesis of AD, which implicates glutamate-induced neuronal damage, this finding reinforces the importance of excitatory neurons in AD pathogenesis64–66. scMAGCA reveals neuronal subtype heterogeneity and regulatory landscapes in Alzheimer’s disease mouse model. a UMAP visualization of cell clusters predicted by scMAGCA, annotated based on known marker genes. b Dot plots showing the average expression levels and proportions of specific marker genes across identified cell types (left), and for marker genes specific to four excitatory neuron subtypes (right). c Bar plot showing the number of cells in each cluster, with orange bars indicating subtypes comprising fewer than 5% of the total cell population. Colors correspond to the subtypes in panel (a). d Heatmap showing GO terms enriched in the four excitatory neuron subtypes. The rows represent enriched GO terms, and the columns correspond to subclusters of excitatory neurons identified by scMAGCA. e The GSEA result shows significant enrichment of the long-term potentiation (LTP) pathway in the EX-1 subtype, consistent with GO term results. f UMAP visualization of the clustering result obtained by scMAGCA on the ATAC modality in the Alzheimer’s disease data, with the annotated cell types consistent with those identified in the RNA modality. g UMAP visualization of gene activity scores for selected cell subgroup-specific marker genes based on scATAC-seq chromatin accessibility in the Alzheimer’s disease dataset. h Heatmap displaying cell type-specific gene activity scores inferred from scATAC-seq chromatin accessibility. Rows represent cell subpopulations and columns represent marker genes. The color scale reflects Z-score-normalized gene activity scores, with warmer colors reflecting higher chromatin accessibility. i Heatmap depicting transcription factor binding motifs specifically enriched in each cell subtype. Darker colors indicate higher enrichment scores, reflecting the regulatory landscape of different cell populations. Utilizing scMAGCA, we further resolved four excitatory neuron subtypes (EX-1 to EX-4) and two inhibitory neuron subtypes. These subtypes exhibited distinct marker gene signatures: Excitatory neurons − 1: EX-1 (Sorcs3, Pdzrn3), Excitatory neurons − 2: EX-2 (Plcl1, Il1rapl2), Excitatory neurons − 3: EX-3 (Rorb, Pld5), and Excitatory neurons − 4: EX-4 (Pik3r3, Dlc1) (Fig. 6b), as well as two inhibitory neuron subtypes: Inhibitory neurons-1 (Gad1) and Inhibitory neurons-2 (Gad2, Lhfpl3). Notably, our findings reveal a higher prevalence of oligodendrocytes, astrocytes, and EX-2 cells, whereas Inhibitory neurons-1 and EX-3 constitute less than 5% of the overall cellular population. Furthermore, pericyte endothelial cells and Type I spiral ganglion neurons are found to be less than 4% of the total (Fig. 6c). In particular, EX-2 and EX-3 displayed strikingly different biological implications in AD. EX-2 (Il1rapl2), which showed an increased proportion in the dataset, is known to facilitate excitatory synapse formation and dendritic spine maturation67, suggesting a potential compensatory role in AD-related synaptic dysfunction. In contrast, EX-3 (Rorb), a well-characterized layer 4 cortical excitatory neuron subtype, has been identified as particularly vulnerable in AD, undergoing progressive degeneration and accumulation of neurofibrillary tangles as the disease progresses68. The decreased proportion of EX-3 neurons in our analysis is consistent with this selective vulnerability, suggesting that EX-3 may represent a neuronal population linked to disease progression. To gain deeper insights into the biological functions of these excitatory neuron subtypes, we performed Gene Ontology (GO) enrichment analysis on their top 100 differentially expressed genes (Fig. 6d). EX-1 neurons were enriched in pathways related to excitatory postsynaptic potential and learning or memory, reinforcing their role in synaptic plasticity69,70. EX-2 neurons were enriched in pathways related to regulation of synapse structure or activity and regulation of postsynaptic membrane potential, further supporting their role in the synaptic organization and maintenance71. Furthermore, to further explore functional differences between EX-1 and EX-2, we conducted Gene Set Enrichment Analysis (GSEA), which revealed a significant enrichment of long-term potentiation (LTP)-associated pathways in EX-1 (Fig. 6e). Since LTP is the molecular basis of learning and memory, our findings suggest that EX-1 neurons play a key role in cognitive processes associated with AD. This neuronal subtype may contribute to synaptic restoration, thereby attenuating the effects of neurodegenerative processes. In addition, these results further validate the ability of scMAGCA to integrate single-cell multi-omics data, enabling precise identification of cell types and their functional attributes through the combined analysis of gene expression and chromatin accessibility. To confirm that scMAGCA effectively integrates matched scRNA-seq and scATAC-seq data, we conducted UMAP visualization using scATAC-seq data alone (Fig. 6f) and computed gene activity scores to infer transcriptional regulation. Genes exhibiting high expression in scRNA-seq also showed elevated gene activity scores in scATAC-seq, reflecting increased chromatin accessibility at regulatory regions. For example, Aqp4 in Astrocyte, Cldn5/Flt1 in Pericytes endothelial, and Slc17a7 in Excitatory neurons displayed strong gene activity enrichment, supporting transcriptional activation at the chromatin level (Fig. 6h). Moreover, cell type-specific marker genes identified in scRNA-seq were likewise specifically enriched in the scATAC-seq data (Fig. 6g). These findings validate that scMAGCA maintains biological consistency across transcriptomic and epigenomic modalities, reinforcing its ability to accurately resolve cell identities. Finally, to further dissect the regulatory mechanisms underlying excitatory neuron diversity in AD, we performed TF motif enrichment analysis in open chromatin regions across different subtypes (Fig. 6i). Our analysis revealed that the four excitatory neuron subtypes exhibit transcription factor binding motifs with varying expression levels, highlighting the diversity of regulatory mechanisms governing these subpopulations. Specifically, EX-1 neurons exhibited strong enrichment for MEF2 family transcription factors (Mef2a/b/c/d) and AP-1 complex-related factors (Fos, Fosb, Bach1/2, Jund), which are known regulators of synaptic plasticity and activity-dependent gene expression. In contrast, EX-2 neurons displayed NeuroD family (Neurod1/4/6) and bHLH factors (Atoh1, Msc) enrichment, which are key drivers of neuronal differentiation and subtype identity maintenance. These results highlight the molecular heterogeneity of excitatory neurons in AD, with EX-1 specialized for experience-dependent plasticity and EX-2 maintaining developmental programs. Given that MEF2 has been implicated in neurodegeneration and synaptic remodeling in AD, its high enrichment in EX-1 neurons suggests a potential role in AD-related excitatory neuron dysfunction72. Emerging evidence shows that MEF2A regulates context-dependent transcriptional responses to inflammatory cues via ERK5 signaling and enhancer accessibility in immune cells, supporting the broader role of MEF2 factors in mediating cellular plasticity, potentially including excitatory neurons under neurodegenerative stress73. Furthermore, in tauopathy mice, deletion of the cGAS DNA sensor increases MEF2C activity in excitatory neurons, restores synaptic structure, and rescues memory; MEF2C targets also overlap with cognition resilience genes in human Alzheimer’s cortex74. Collectively, these findings position MEF2-driven enhancer regulation as a core mechanism of neuronal plasticity under neurodegenerative stress, providing mechanistic support for the strong MEF2 motif enrichment observed in our EX-1 neurons. Together with the broader integrative single-cell analysis provided by scMAGCA, this supports a refined, multi-omic view of excitatory-neuron regulation in Alzheimer’s disease. scMAGCA can identify disease-associated cell subpopulations from joint modalities and infer potential biomarkers for kidney cancer Kidney cancer encompasses a diverse range of cellular states and molecular alterations, complicating both diagnosis and treatment. Single-modality profiling often fails to capture the full extent of heterogeneity, especially when distinguishing tumor-promoting cells from non-malignant counterparts. By jointly analyzing transcriptomic and chromatin accessibility data, scMAGCA provides a multi-layered view of tumor biology, enabling the detection of subtle disease-associated cell subpopulations and revealing biomarkers linked to cancer progression. In our study, we applied scMAGCA to the Mouse_kidney dataset, which comprises transcriptomic and chromatin accessibility profiles from 14,652 single cells. First, we performed scMAGCA on this dataset and obtained ten clusters. These clusters are annotated into seven predominant cell types associated with kidney cancer by visually representing the expression patterns of selected marker genes (Fig. 7a and Supplementary Fig. 36). Notably, our approach resolved two disease-associated endothelial subtypes and three proximal tubule subtypes, demonstrating its ability to detect fine-grained cellular heterogeneity. While clustering based on RNA data alone revealed some heterogeneity, multi-omics integration (RNA+ATAC) enhanced cell-type resolution. In contrast, clustering using ATAC-only data or Seurat yielded less distinguishable clusters (Supplementary Figs. 37, 38). Moreover, we assess the clustering and integration quality of scMAGCA using ASW, DB, and CH, and compare its performance with established multimodal integration frameworks, including GLUE23, Seurat28, and MultiVI45. Details of this analysis are provided in Supplementary Note 10. scMAGCA efficiently identifies disease-associated subtypes as well as cancerous and non-cancerous cell types from the kidney multi-omics data. a UMAP visualization of scMAGCA clustering results on the multi-omics kidney data using marker genes, including three subtypes in proximal tubule cells and two subtypes in endothelial cells. b Heatmap showing the RNA expression (left) and chromatin accessibility (right) of ten renal cell types. c UMAP visualization of the scMAGCA clustering result to identify the cancerous and non-cancerous cell types based on copy number data and the proportion of aneuploid, diploid, and undefined copy number states across different renal cell types. d Volcano plot showing the differentially expressed genes in upregulated, stable, and downregulated genes from endothelial cell 1 and endothelial cell 2. Differential expression was assessed with the default two-sided Wilcoxon rank-sum test. p-values were adjusted for multiple comparisons using the Benjamini-Hochberg false discovery rate method. e Functional enrichment analysis of upregulated and downregulated differentially expressed genes. The x-axis shows the number of genes annotated to each term; bubble size is proportional to gene count, and bubble color represents the enrichment significance as −log10(p). The one-sided hypergeometric test was used for enrichment analysis, and p-values were adjusted for multiple testing using the Benjamini-Hochberg false discovery rate method. f The number and percentage of genes co-regulated by RNA and ATAC genomics in endothelial cell 1 and endothelial cell 2. g Chromatin accessibility profiles surrounding the Hnf4a genomic region, with corresponding gene expression distribution across renal cell populations shown in violin plots. Peak-to-gene connections are indicated by loops beneath the accessibility track, while the bottom panel presents ZNF263 motif enrichment analysis within the peak region. h TF footprint for the ZNF263 motif using Tn5 bias-subtraction normalization. i Differential gene expression patterns and genomic tracks for selected genes in two endothelial cells. Violin plots show expression levels of Eng, Meis2, Gramd1b, and Slco1a6, with corresponding genomic browser views displaying read coverage in their chromosomal regions. To further validate these annotations, we performed differential expression and chromatin accessibility analysis, revealing cell-type-specific transcriptional signatures. Notably, ATAC peaks aligned with corresponding transcriptomic expression patterns (Fig. 7b), reinforcing the utility of integrating chromatin accessibility data for cell identity determination. We next applied CopyKAT to distinguish cancerous from non-cancerous cells, revealing a clear segregation (Fig. 7c). Malignant populations were predominantly enriched in endothelial cell 1 and proximal tubule cells 1 and 2, indicating that these subtypes represent potential tumorigenic lineages. To explore the molecular drivers of these cancer-associated subtypes, we investigated the cancer marker gene Hnf4a75. Hnf4a expression and chromatin accessibility were elevated in endothelial cell 1 and proximal tubule cells 1 and 2 (Fig. 7g), further supporting their malignant identity. Furthermore, motif analysis using Signac’s FindMotifs function identified ZNF263 as a putative regulator of Hnf4a transcription (Fig. 7g). ZNF263, a zinc finger transcription factor, has been implicated in kidney-related gene regulation and may influence tumor progression. Transcription factor footprinting analysis revealed stronger ZNF263 binding in proximal tubule cells, suggesting a potential regulatory role in these cancer subtypes (Fig. 7h). To further substantiate the functional relevance of the regulatory programs identified by scMAGCA, we experimentally validated the role of HNF4A, which our model implicates as a key regulator in the kidney tumor context. Specifically, we used siRNA-mediated knockdown of HNF4A in 786-O and Caki-1 kidney cancer cell lines and assessed the effects on cell proliferation and migration. CCK-8 assays showed that HNF4A knockdown markedly reduced the viability of 786-O and Caki-1 cells over time, and colony formation assays demonstrated a decrease in clonogenic growth compared with control cells (Supplementary Fig. 40a–d). Transwell migration assays further revealed that siHNF4A treatment significantly impaired the migratory capacity of both 786-O and Caki-1 cells (Supplementary Fig. 40e, f). These quantitative phenotypic measurements indicate that HNF4A promotes the proliferation and migration of kidney cancer cells, providing independent experimental support for the pro-tumorigenic role assigned to HNF4A by the scMAGCA model in the kidney analysis. To characterize cancer-associated transcriptional and epigenetic alterations, we performed differential expression analysis comparing cancerous and non-cancerous subtypes of Endothelial and proximal tubule cells. In endothelial cells, upregulated genes were primarily associated with metabolic reprogramming and mitochondrial function, whereas downregulated genes indicated impairments in vascular integrity and endothelial homeostasis (Fig. 7d, e). By integrating ATAC-seq data, we identified 360 genes exhibiting increased expression in regions of open chromatin and 38 genes with reduced expression in regions of closed chromatin, suggesting chromatin remodeling associated with tumorigenesis (Fig. 7f). Among these, Eng and Meis2-both known to drive tumor angiogenesis-were markedly upregulated. In contrast, Gramd1b and Slco1a6 were downregulated in lipid metabolism and solute transport. These gene expression changes were accompanied by corresponding shifts in chromatin accessibility, further reinforcing their functional relevance (Fig. 7i). Among proximal tubule cells, subtypes 1 and 2 exhibited cancer-associated transcriptional signatures, suggesting their involvement in tumor progression. We identified 137 genes with increased expression in open chromatin regions and 105 genes with reduced expression in closed chromatin regions, consistent with epigenetic remodeling in tumorigenic cells (Fig. 8a). Notably, Slc34a1 and Snx29 displayed pronounced upregulation, suggesting their involvement in tumor-associated metabolic pathways. Conversely, Pparg and Car3 were markedly downregulated, indicating potential disruptions in lipid metabolism and oxidative stress regulation (Fig. 8b). Given their roles in metabolic reprogramming and cellular stress response, Slc34a1, Snx29, Pparg, and Car3 emerge as potential therapeutic targets, warranting further investigation. scMAGCA discovers the potential biomarkers on the dynamic progression of kidney cancer and finds that LACTB2 and NCOA2 may act as pivotal regulators in the oncogenesis of proximal tubule cells. a The number and percentage of genes categorized by RNA expression and ATAC accessibility in proximal tubule cells. b Distinct gene expression profiles and genomic landscapes for candidate genes across three proximal tubule cell subtypes. Violin plots illustrate the expression distribution of Pparg, Car3, Slc34a1, and Snx29, alongside genomic browser snapshots showing read distribution across their respective chromosomal loci. c Pseudotime trajectories of proximal tubule cell clusters plotted on the first two principal components. Relative expression of key genes (Adhfe1, Cspp1, Lactb2, Lypla1, Ncoa2) across pseudotime is shown. For expression trends, solid lines represent LOESS-fitted mean values, and shaded bands indicate 95% confidence intervals of the fit. d Gene expression and chromatin activity scores of selected genes across proximal tubule cell populations (n = 15,425 independent cells). Significance was assessed using a two-sided Welch’s t test without multiple-comparison adjustment, comparing cancer-associated (Proximal tubule cells 1 and 2) and normal (Proximal tubule cell 3) populations. Exact p-values, test statistics, degrees of freedom, effect sizes and confidence intervals are provided in Supplementary Table 15 (ns: p > 0.05, *: p < 0.05, **: p < 0.01; ***: p < 0.001). e Cell interaction network illustrating the communication between proximal tubule cells and other cell types identified by scMAGCA within the renal tubular system. f GAS and EGF signaling pathways illustrate predicted intercellular communication among seven cell clusters. Potential signaling interactions are represented by links between source clusters (filled circles, indicating high ligand gene expression) and target clusters (unfilled circles, indicating high receptor gene expression). Each circle’s color denotes a distinct cell cluster. g Using quantitative real-time polymerase chain reaction (qPCR) to examine the RNA expression levels of LYPLA1, LACTB2, NCOA2, ADHFE1, and CSPP1 in kidney cancer cell lines relative to the normal renal proximal tubule cell line HK-2. Graph shows mean ± s.d. from n = 3 biologically independent experiments for each cell line. Each point represents one biologically independent experiment. The HK-2 cell line served as the control group. Statistical significance was assessed using a two-sided t test without multiple-comparison adjustment. Exact p-values, test statistics, degrees of freedom, effect sizes and confidence intervals are provided in Supplementary Table 16. h Quantitative real-time PCR analysis of LACTB2 and NCOA2 in paired kidney cancer and adjacent normal tissues. Data are presented as mean ± SD with individual data points overlaid; n = 7 paired kidney cancer tissues and adjacent normal kidney tissues. Adjacent normal tissues from the same patients served as controls. Statistical significance was assessed across paired kidney cancer and adjacent normal tissues using a two-sided t test without multiple-comparison adjustment. Exact p-values, test statistics, degrees of freedom, effect sizes and 95% confidence intervals are provided in Supplementary Table 17. To investigate cell-cell communication in tumor progression, we employed CellChat76 to analyze signaling interactions between proximal tubule cells and their microenvironment. Proximal tubule cell 1 exhibited strong intercellular signaling, both in autocrine (self-signaling) and paracrine (cross-cell) interactions, suggesting a key role in the tumor microenvironment (Fig. 8e). Notably, a progressive communication trajectory was observed, with normal proximal tubule cell 3 signaling to early cancerous proximal tubule cell 2, which then relayed signals to advanced cancerous proximal tubule cell 1 (Supplementary Fig. 41). This hierarchical signaling cascade suggests a coordinated tumorigenic transition among proximal tubule cells. We identified several critical signaling pathways, including BMP, SPP1, EGF, FGF, and GAS (Supplementary Fig. 42). Among these, GAS and EGF signaling emerged as central regulators of tumor progression (Fig. 8f). GAS signaling was strongly linked to immune evasion, potentially allowing cancer cells to circumvent immune-mediated destruction77. Meanwhile, EGF activation was associated with an enhanced proliferative capacity and metastatic potential78, reinforcing its role as a key driver of kidney cancer progression. These pathways are crucial in kidney function and pathology, suggesting potential therapeutic targets. To elucidate the dynamic transitions in proximal tubule cell states, we performed pseudotime analysis, reconstructing the progression from normal cells (Proximal tubule cell 3) to early-stage cancer (Proximal tubule cell 2) and ultimately to advanced cancer (Proximal tubule cell 1) (Fig. 8c). Functional enrichment analysis (Supplementary Fig. 43) demonstrated distinct roles across this trajectory. Specifically, normal proximal tubule cells (Proximal tubule cell 3) exhibited enrichment for AMPK signaling and organelle homeostasis terms, consistent with a homeostatic program. While cancerous cells (Proximal tubule cells 1 and 2) were enriched for broad metabolic and transport processes as well as MAPK signaling and transcriptional activation, indicating activated and proliferative programs. Pseudotime analysis revealed transcriptional changes along the tumorigenic trajectory, with Adhfe1, Cspp1, Lactb2, Lypla1, and Ncoa2 showing statistically significant differences in gene expression in cancer-associated proximal tubule cells 1 and 2 compared to normal proximal tubule cells (Fig. 8d and Supplementary Table 15). Among these, Lactb2, a gene that enables RNA endonuclease activity, plays a pivotal role in mitochondrial homeostasis and cell viability79, while Ncoa2, a nuclear receptor coactivator, is essential for tumor cell proliferation and differentiation80. The alterations of these genes suggest their roles in kidney cancer progression. Using the scMAGCA latent factor representation to quantify gene-factor associations, we found that, as shown in Supplementary Fig. 44a, NCOA2 and LACTB2 are both among the top-loading genes of Factor 25 in the cancer-associated Proximal tubule cell 2 population of the kidney dataset. GO and KEGG enrichment (Supplementary Fig. 44b, c) of Factor 25-associated genes in cancer-associated Proximal tubule cell 2 indicates catabolic and stress-adaptive responses, enhanced migratory potential, and cancer-related transcriptional and extracellular matrix remodeling programs. To assess the clinical relevance of these findings, we analyzed the expression of their human homologs in a diverse panel of kidney cancer cell lines using quantitative polymerase chain reaction (qPCR). Consistently, the qPCR analysis revealed the significantly elevated expression of these genes in kidney cancer cell lines relative to the normal renal proximal tubule cell line HK-2 (Fig. 8g and Supplementary Table 16). Furthermore, the expression of LACTB2 and NCOA2 was significantly up-regulated in patient-derived kidney cancer tissues compared to the adjacent normal renal tissues (Fig. 8h and Supplementary Table 17), further supporting their role in tumorigenesis. Together, these findings highlight LACTB2 and NCOA2 as potential oncogenic drivers in kidney cancer, demonstrating the utility of scMAGCA in identifying clinically relevant biomarkers and uncovering key molecular mechanisms of renal carcinogenesis. Discussion In this study, we introduce scMAGCA, a single-cell multi-omics deep learning framework designed to address the intrinsic complexity and noise inherent in single-cell multi-omics (scMulti-omics) data. By leveraging a graph adversarial learning-based multi-modal autoencoder, scMAGCA effectively integrates RNA and ADT/ATAC data, enabling reliable cross-modality alignment while preserving cellular topology. It bridges distinct omics layers through an unsupervised cell-graph, maintaining flexibility and broad applicability without sacrificing the power of graph-based feature propagation. Specifically, scMAGCA reconstructs latent representations through a dual-decoder architecture, comprising a linear decoder and a Zero-Inflated Negative Binomial (ZINB) decoder, which collaboratively reconstructs the count matrix, ensuring accurate data representation. Furthermore, scMAGCA employs an adversarial learning module, aligning the distribution of the latent embedding with the prior distribution to enhance embedding quality. During the clustering stage, scMAGCA utilizes Kullback-Leibler (KL) divergence to refine cell-type separation, enabling the effective resolution of heterogeneous cellular populations. By integrating these components, scMAGCA not only enhances data harmonization across omics layers but also improves cell-type identification and biological interpretability in complex multi-omics datasets. To validate the effectiveness of scMAGCA in scMulti-omics datasets, including RNA+ADT and RNA+ATAC datasets, we compared its performance against multiple multi-modal clustering methods. Experimental results demonstrate that scMAGCA efficiently characterizes heterogeneous count data from scMulti-omics datasets, surpassing alternative approaches in both accuracy and scalability. Notably, scMAGCA effectively identified subpopulations that remain indistinguishable using conventional methods while preserving their intrinsic biological relationships. Moreover, scMAGCA outperformed multiple state-of-the-art methods in batch effect removal on the multi-batch datasets, highlighting its ability to integrate single-cell modalities across batches while maintaining biologically interpretable representations. Beyond clustering, scMAGCA generates interpretable multi-omics embeddings that uncover functional genomic relationships while accurately distinguishing subtle cellular subpopulations undetectable by traditional approaches. This dual ability to maintain fine-grained resolution while ensuring biological coherence helps to identify cellular states and their underlying regulatory mechanisms. The clustering outcomes generated by scMAGCA are critical for downstream analyses in multi-omics studies. By simultaneously integrating transcriptomic profiles, surface protein expression, and chromatin accessibility patterns, scMAGCA delineates B cell subpopulations with distinct chromatin accessibility states that correlate with immune regulatory pathways. This tri-modal approach enables a more comprehensive resolution of cell identity, overcoming limitations inherent in single or dual-modality analyses. In Alzheimer’s disease, scMAGCA uncovered distinct transcription factor binding motifs and regulatory mechanisms associated with each excitatory neuron subtype, particularly between EX-1 and EX-2 neuronal populations. These differences were marked by the enrichment of MEF2 and NeuroD family transcription factors, underscoring their distinct roles in synaptic plasticity and neuronal development. In kidney cancer, scMAGCA identified cell-type-specific peak-gene associations and characterized transcriptional patterns across diverse cell populations, leading to the discovery of biomarkers. Notably, qPCR validation confirmed that LACTB2 and NCOA2 exhibited significantly altered expression levels in kidney cancer cell lines compared to the normal renal proximal tubule cell line (HK-2), as well as in kidney cancer patient samples compared to matched adjacent normal renal tissues. These findings highlight their potential roles in kidney cancer progression, reinforcing the utility of scMAGCA in biomarker discovery and disease characterization. In summary, scMAGCA provides an interpretable framework for integrative analyses of single-cell multi-omics data, demonstrating its potential to advance deep learning applications in single-cell biology. Looking ahead, we aim to extend scMAGCA to accommodate emerging multi-omics modalities and ultra-large datasets, further enhancing its scalability, analytical capabilities, and computational efficiency; in parallel, we will pursue comprehensive optimization of cross-modal prediction (RNA → ADT/ATAC) as a meaningful direction for future work81. Methods Ethical approval statement The research protocol for the human tissue sample validation of this study had been approved by the Ethics Committee of Peking University Third Hospital (Approval No.: IRB00006761-M20260071) and was conducted in accordance with the principles of the Declaration of Helsinki. The human datasets involved for bioinformatics analysis were all from publicly available databases with anonymized patient information. Human kidney cancer tissues The human kidney cancer tissues and paired paracancerous tissues (n = 7 per group) used in this study were obtained from patients enrolled at the Department of Urology, Peking University Third Hospital, Beijing, China. The study was carried out in accordance with the Declaration of Helsinki and was approved by the Ethics Committee of Peking University Third Hospital (Approval No.: IRB00006761-M20260071). All participants provided written informed consent prior to enrollment, and participation was voluntary without any financial compensation. For the analysis in Fig. 8h, the study included seven patients, comprising four males and three females (sex determined based on patient self-report). Age is reported as a ranges rather than exact values, to reduce indirect identification risk. Detailed clinical characteristics, including patient sex, age range, pathological type, tumor stage and tumor grade, are provided as individual-level data in the Source Data file. The analysis for Fig. 8h did not account for gender, as there was no evidence that gender significantly affected the expression levels of the genes of interest in this study. Furthermore, the distribution of four males and three females was relatively comparable. Data harmonization and normalization Single-cell multi-omics data are highly heterogeneous, and count-based modalities such as scRNA-seq and scATAC-seq are markedly zero-inflated and sparse, posing substantial challenges for integrative analysis while preserving key biological signals. To meet these challenges, we apply a modality-specific preprocessing workflow that harmonizes data distributions while retaining information essential for downstream modeling. For RNA+ADT datasets, preprocessing is performed with the SCANPY package13. Each modality undergoes three steps-(i) feature selection, (ii) library-size normalization, and (iii) loge(x+1) transformation. Features are ranked jointly by dropout rate and mean log-expression; those that are either highly expressed and widely detected, or moderately expressed but consistently present, are retained. Concretely, we keep the top 2000 highly variable genes for the RNA modality, while all Antibody-Derived Tags (ADTs) are kept because they are few in number and exhibit low sparsity. The RNA+ATAC datasets are processed identically, selecting the top 2000 highly variable genes and the most variable ATAC peaks. TEA-seq datasets follow the same scheme for genes and peaks, and again, all ADTs are retained. Afterward, cell counts are normalized to correct for technical variability, followed by a normalization and log-transform to enforce zero mean and unit variance. We construct two types of concatenated feature matrices, namely the preprocessed matrix Xcon and the raw-input matrix after feature selection Xrawcon. Specifically, Xcon is constructed by horizontally concatenating the matrices of each modality along the column dimension. For instance, in RNA+ADT datasets, Xcon∈Rn×m, where m = dadt + drna, is formed by concatenating Xd∈Rn×dadt and Xr∈Rn×drna, where n denotes the number of cells and dadt/drna are the numbers of features in each modality. This processed matrix Xcon retains informative features and serves as the input for downstream modules such as graph construction and the linear reconstruction decoder. In contrast, Xrawcon∈Rn×m is constructed by concatenating the raw modality-specific matrices after feature selection without any transformations. Due to its high sparsity and zero inflation, we use Xrawcon only as input to the ZINB loss. This allows the model to accurately capture dropout patterns and better reflect the true underlying distribution of the original multi-omics data31. To ensure clarity and avoid confusion in notation, we have provided a detailed notation table in Supplementary Note 11 (Supplementary Table 18). Multi-omics graph construction To preserve cell-cell relationships across different omics layers, we construct a multi-omics cell-cell graph, unifying RNA, protein, and chromatin accessibility signals into a single structured representation. We apply the K-nearest neighbors (KNN) algorithm34 on Xcon to define an adjacency matrix A, where each node represents a cell, edges indicate cell-cell similarity, and cosine similarity is used to quantify intercellular distance. Concretely, for the i-th and j-th cells in Xcon: where Xicon and Xjcon denote the i-th and j-th rows of Xcon, respectively. Based on the k-nearest neighbors, the adjacency matrix A is then defined as: where Nk(i) represents the neighborhood that contains the nearest k nodes with the i-th node as the center. The final multi-omics cell-cell graph is then represented as: where G encodes both the cell-feature matrix (Xcon) and the adjacency matrix (A). In this way, we unify the multi-omics signals within a single topological structure, allowing subsequent modules in scMAGCA to leverage local cell-cell relationships for robust embedding and clustering. Multi-omics adversarial graph convolutional autoencoder High-dimensional, heterogeneous single-cell multi-omics data requires a strategy that maintains local topological relationships while robustly modeling dropout and other technical noise. Simply reducing dimensionality using a conventional autoencoder often loses neighborhood structure and cannot properly handle abundant zero counts. To accurately capture cellular graph architecture and inter-node relationships in single-cell multi-omics data, we propose an adversarial graph convolutional autoencoder (AGC-AE), which unifies graph-based feature extraction, ZINB-based modeling, and adversarial alignment into a unified framework. To encode the cell-cell relationships captured in the multi-omics graph G(Xcon,A), AGC-AE employs a graph convolutional encoder. The encoder projects each cell’s multi-omics profile into a lower-dimensional latent representation Z, while preserving its connectivity and neighborhood structure. The encoding operation for the l-th GCN layer is formulated as: where Z(l) represents the feature embeddings at layer l, and Z(l+1) is the output after convolution. Specifically, the input layer is defined as Z0=Xcon∈Rn×m. The layer-wise update function is formulated as: where A~=A+I,D~ii=∑jA~ij, We is the learnable parameter matrix in the neural network, I is the identity matrix of A, and σ(⋅) is a PRelu activation function. The final encoder output Z is a latent representation capturing both the multi-omics signals and local graph structure. After embedding, AGC-AE reconstructs the integrated multi-omics matrix Xcon using both a linear decoder and a ZINB decoder. The linear decoder ensures the global fidelity of multi-omics features by minimizing the mean squared error (MSE) loss: where Xcon′ represents the reconstructed cell matrix from the linear decoder and Wlinear represents the learnable weight matrix of the decoder. Meanwhile, AGC-AE employs the ZINB model as the loss function for all scMulti-omics data, using the feature-selected raw data matrix as input31,41. To model data characteristics, the ZINB model is designed to accommodate the zero-inflated nature of single-cell data by introducing three key parameters: the mean (μ) of the negative binomial distribution to capture central tendencies, the dispersion (θ) to quantify variability, and an additional coefficient (π) that represents the probability of the likelihood of zero-values or dropout events: To ensure accurate estimation of these parameters, AGC-AE integrates three independent fully connected layers within the latent space Z, each responsible for predicting μ, θ, and π, which can be formulated as follows: where Wμ, Wθ, and Wπ are the learnable weight matrices associated with each parameter. The activation functions σμ(⋅) and σθ(⋅) are of the exponential type because the mean and dispersion parameters are non-negative, while σπ(⋅) uses a sigmoid activation to constrain the dropout probability within [0,1]. All these three matrices, the mean matrix M, dispersion matrix Θ, and dropout-probability matrix Π, have dimension Rn×m, where n is the number of cells and m is the total number of features across all input modalities after feature selection (for example, drna RNA + dadt ADT features for the RNA+ADT dataset). These parameters are used to compute the ZINB likelihood for each entry in the combined raw feature matrix after feature selection Xrawcon. The final ZINB-based loss function used in AGC-AE is then formulated as: By integrating ZINB loss within the AGC-AE framework, scMAGCA effectively disentangles true biological variation from technical artifacts, ensuring that downstream clustering and regulatory inference retain a biologically meaningful structure. This enables the model to leverage both the graph-based topological consistency and zero-inflated modeling of sparse data, thus improving robustness in multi-omics cell type identification and feature extraction. Modality alignment via adversarial learning Multi-omics data often exhibits modality-specific distributional shifts, where distinct omics layers are embedded into latent spaces with varying degrees of separation. Standard alignment techniques, such as Maximum Mean Discrepancy (MMD)82,83 and Siamese networks84, often fail to fully mitigate these differences, leading to suboptimal cross-modal integration. To address this, scMAGCA employs an adversarial alignment strategy, where a generator (G) and a discriminator (Ddisc) engage in a minimax game to enforce modality-invariant representations. This approach allows the latent embeddings of different omics modalities to be projected into a common structured space while retaining their complementary information. Within this adversarial framework, the generator (G) is served by the graph encoder, while the discriminator (Ddisc) is used to distinguish whether a latent code z is from the prior p(z) (positive) or from the generator (negative). Specifically, the adversarial optimization aims to minimize the discrepancy between real and generated latent distributions by iteratively updating both networks: Here, we set the latent prior p(z)=N(0,IdZ), where dZ is the latent dimensionality, as is commonly done in adversarial alignment85,86, which helps to center and scale the embeddings. In the implementation, the output of the discriminator is a scalar score, which evaluates whether a given latent embedding belongs to the prior distribution or has been generated by the encoder. Cells sampled from the prior distribution are treated as real (label = 1), whereas encoder-derived latent representations are treated as generated samples (label = 0). The final adversarial loss is defined as: where zip,zi, and n represent the i-th sample generated from the prior distribution p, the latent representation of the i-th cell, and the number of cells, respectively. By integrating adversarial learning into our modality-aware latent alignment, scMAGCA effectively bridges the distributional gap between omics layers while maintaining biologically meaningful variation. Unlike traditional methods that rely solely on statistical similarity measures, in scMAGCA, the adversarial alignment is applied to the joint latent space learned by the GCN-augmented encoder, synchronizing the modality integration and alignment in one cohesive model. K-means clustering with KL divergence refinement To enhance clustering performance by reinforcing associations between similar cells in the latent space, we introduce a two-stage clustering strategy that first applies K-means initialization and then refines cluster assignments using Student’s t-distribution similarity and KL divergence minimization. First, given a latent representation zi and a cluster centroid μj, we compute the soft assignment probability qij using Student’s t-distribution, ensuring that cells closer to a centroid receive a higher similarity score: where qij represents the soft assignment of zi to cluster μj. To further refine cluster assignment, we construct a target distribution (P) that enhances high-confidence assignments while suppressing weak associations. The target distribution (P) is defined as: where P amplifies confident cluster assignments by squaring Q and normalizing over all clusters. To further optimize the clustering process, we construct an embedded clustering loss function Lc to minimize the KL divergence between the target distribution P and the initial assignment matrix Q, iteratively refining cell-cluster assignments: End-to-end multi-omics embedding and clustering strategy To achieve an end-to-end optimization for single-cell multi-omics clustering, we formulate a two-stage training paradigm that integrates adversarial representation learning, zero-inflated modeling, and clustering refinement into a unified framework. The first stage, pretraining, ensures that the latent embedding effectively reconstructs multi-modal data while aligning different omics distributions. The second stage, clustering refinement, initializes cell clusters and iteratively enhances their separation via KL divergence minimization. Pretraining for multi-omics representation learning: During the pretraining phase, the model optimizes an objective function that balances three key loss components: (1) Reconstruction fidelity using Mean Squared Error (MSE) for preserving global multi-omics structure. (2) Dropout-aware modeling via ZINB to capture sparsity in single-cell omics. (3) Modality alignment enforced by adversarial learning to ensure inter-omics consistency. Therefore, the total pretraining loss is thus formulated as: where α, β, and γ control the relative contribution of each loss term. Unsupervised clustering refinement: After pretraining, K-means clustering is applied to initialize cluster centers in the learned latent space. To improve the separation of cell states, we introduce KL divergence refinement, where cluster assignments are iteratively refined to maximize high-confidence cell-cluster associations while suppressing uncertain assignments. The final training loss incorporates clustering refinement as follows: To optimize the graph convolutional autoencoder and clustering assignments, we employ stochastic gradient descent (SGD) with backpropagation. The gradients for each latent cell representation zi and cluster center μj are updated as: where pij and qij represent the target and predicted cluster assignments, respectively. For a dataset with n cells and learning rate lr, the clustering center μj is updated by: Optimization and parameter updates: In the pretraining stage, the encoder weights We are updated by: Then, in the training stage, the encoder weights We are updated by: Finally, the linear decoder weight Wlinear is updated by: Functional genomics interpretation from integrated multi-omics representations In scMAGCA, the latent embedded representations are generated using a graph convolutional autoencoder with four GCN layers. Through this multi-stage transformation, the combined cell-gene matrix Xcon∈Rn×m is projected into a 32-dimensional latent space. For instance, in RNA+ATAC datasets, m corresponds to 2000 genes and 2000 peaks, while in RNA+ADT datasets, it consists of 2000 genes and the number of ADTs. The weight matrix (We) of each GCN layer is central to the model’s representational capacity, playing a critical role in dimensionality reduction and feature learning. To interpret biological relevance, we analyze the relationship between the weight matrices of GCN layers and the latent representations (Z). By tracing backward from the latent space to the input layer, we identify key genes, peaks, or ADTs that contribute to the learned representations. First, the output of the first GCN layer, Z1∈Rn×1024, is computed from the input matrix Z0=Xcon∈Rn×m and the first-layer weight matrix We0∈R1024×m using Equations (3) and (4). Similarly, the final latent representation (Z∈Rn×32) is derived from the third GCN layer (Z3∈Rn×64) using the weight matrix (We3∈R32×64). Thus, the learned latent representation Z encodes integrated information from genes, peaks, or ADTs, which can be attributed to weights in We3. To analyze the contribution of each weight, the standard deviation of We3 is calculated, yielding a result vector r3∈R1×64. This vector is then sorted, and the indices of the top 50% (i.e., 64 × 0.5 = 32 entries) are selected, forming the index vector ind2∈R32×1. Using these indices, the columns of We2∈R64×256 corresponding to the top-ranked weights are retained, resulting in the reduced matrix W¯e2∈R32×256. This process is iteratively applied, layer by layer, progressing outward. From the first-layer weight matrix We0∈R1024×m, a reduced matrix W¯e0∈R512×m is obtained by retaining the top-ranked weights based on ind0∈R512×1. Finally, the indices of the top 200 entries, represented as ind∈R1×200, are extracted. These indices correspond to 200 key features, a combination of genes and peaks or ADTs, which are identified and biologically interpreted from the original input matrix Xcon∈Rn×m. Moreover, Supplementary Fig. 48 demonstrates that even with non-linear activations and multi-hop message passing, most high-priority features remain traceable from middle layers back to the input. This observation confirms that the attribution signal is preserved across network depths. Our backward-tracing approach thus provides an effective way to interpret GCN-learned embeddings in multi-omics settings, balancing structural abstraction with feature-level interpretability. Model implementation For each multi-omics dataset, scMAGCA uses a graph convolutional autoencoder with a fully connected encoder of layer widths {m, 1024, 256, 64, 32}, where m is the number of features in the concatenated input matrix, and a decoder with layers {32, 512, 1024, m} that reconstructs the original feature space. In all experiments, we construct the KNN-based adjacency using k = 30 nearest neighbors per cell, and k is fixed globally across all datasets and modalities. The 32-dimensional bottleneck layer defines the latent representation. During the pretraining stage, we minimize a composite loss consisting of a weighted reconstruction loss, a ZINB loss, and an adversarial loss, together with a KL-divergence regularization term. Then, the weights for the reconstruction, ZINB, and adversarial components are fixed to α = 0.2, β = 0.8, and γ = 0.01, respectively. Pretraining is performed for 400 epochs using the Adam optimizer with AMSGrad and an initial learning rate of 1.0 × 10−3. After pretraining, we initialize cluster centroids by applying K-means to the latent embeddings and then enter the clustering stage, in which the latent representation and cluster centroids are jointly refined by continued optimization of the same composite loss. We use the scikit-learn KMeans implementation (from sklearn.cluster import KMeans). In our code, we set all arguments to scikit-learn defaults (init = ’k-means ++’, n_init = 10, max_iter = 300, tol = 1e-4). We additionally use random_state = None (scikit-learn default) and therefore draws randomness from NumPy’s global RNG; we set np.random.seed(seed) at the beginning of each run, making the K-means initialization reproducible given the same seed. The clustering stage uses the Adadelta optimizer with a learning rate of 1.0 and ρ = 0.95, is run for at most 2000 iterations, and is considered converged when fewer than 0.1% of cell labels change between consecutive epochs. For all baseline methods, we adopt the authors’ recommended or default parameter settings or, where appropriate, the tuning strategies described in the original papers as analyzed in Supplementary Note 12. We run the entire pipeline 10 independent times per dataset, and report mean ± standard deviation for all clustering metrics. Competing methods scMDC31 (https://github.com/xianglin226/scMDC). It is an integrated deep multi-modal autoencoder, which uses an encoder to fuse diverse modal data and twin decoders to restore each modality individually. scMDC clearly distinguishes data origins and jointly learns hidden embeddings for clustering insights. scMDC31 (https://github.com/xianglin226/scMDC). It is an integrated deep multi-modal autoencoder, which uses an encoder to fuse diverse modal data and twin decoders to restore each modality individually. scMDC clearly distinguishes data origins and jointly learns hidden embeddings for clustering insights. DeepMAPS34 (https://github.com/OSU-BMBL/deepmaps). It is a tool for inferring biological networks from scMulti-omics data. DeepMAPS represents scMulti-omics as a heterogeneous graph and employs a multi-head graph transformer to learn cell-gene interactions in localized and broad contexts robustly. DeepMAPS34 (https://github.com/OSU-BMBL/deepmaps). It is a tool for inferring biological networks from scMulti-omics data. DeepMAPS represents scMulti-omics as a heterogeneous graph and employs a multi-head graph transformer to learn cell-gene interactions in localized and broad contexts robustly. BREMSC19 (https://github.com/tarot0410/BREMSC). It is a Bayesian Random Effects Mixture model integrating MCMC and Dirichlet distributions, designed to cluster paired single-cell CITE-seq data simultaneously. BREMSC19 (https://github.com/tarot0410/BREMSC). It is a Bayesian Random Effects Mixture model integrating MCMC and Dirichlet distributions, designed to cluster paired single-cell CITE-seq data simultaneously. totalVI20 (https://github.com/YosefLab/totalVI_reproducibility). It is a deep generative tool for CITE-seq data, learning a joint probability representation that considers unique noise, biases, and batch effects from each data modality. totalVI20 (https://github.com/YosefLab/totalVI_reproducibility). It is a deep generative tool for CITE-seq data, learning a joint probability representation that considers unique noise, biases, and batch effects from each data modality. MOFA +33 (https://github.com/bioFAM/MOFA2). It is a Bayesian group factor analysis framework to integrate multi-omics data, inferring a compact representation through a few latent factors that capture key variability sources. MOFA +33 (https://github.com/bioFAM/MOFA2). It is a Bayesian group factor analysis framework to integrate multi-omics data, inferring a compact representation through a few latent factors that capture key variability sources. Seurat (V4)28 (https://github.com/satijalab/seurat). It is an R toolset for analyzing multi-omics data, using an unsupervised approach with weighted nearest neighbors to determine modality weights unique to each cell. This enhances the integration of diverse data types in analysis. Seurat (V4)28 (https://github.com/satijalab/seurat). It is an R toolset for analyzing multi-omics data, using an unsupervised approach with weighted nearest neighbors to determine modality weights unique to each cell. This enhances the integration of diverse data types in analysis. Cobolt21 (https://github.com/epurdom/cobolt). It is a multi-modal variational autoencoder based on a hierarchical Bayesian generative model, which is designed to analyze and integrate multi-modality datasets with single-modality data. Cobolt21 (https://github.com/epurdom/cobolt). It is a multi-modal variational autoencoder based on a hierarchical Bayesian generative model, which is designed to analyze and integrate multi-modality datasets with single-modality data. scDeepCluster41 (https://github.com/ttgump/scDeepCluster). It is a deep zero-inflated negative binomial (ZINB)-based clustering method designed for scRNA-seq data, which efficiently maps read count matrices into low-dimensional representations. It can be extended to multi-omics data. scDeepCluster41 (https://github.com/ttgump/scDeepCluster). It is a deep zero-inflated negative binomial (ZINB)-based clustering method designed for scRNA-seq data, which efficiently maps read count matrices into low-dimensional representations. It can be extended to multi-omics data. GLUE23 (https://github.com/gao-lab/GLUE). It is a computational framework based on graph-guided variational autoencoders, designed for integrating single-cell multi-omics data and inferring cross-omics regulatory relationships. GLUE23 (https://github.com/gao-lab/GLUE). It is a computational framework based on graph-guided variational autoencoders, designed for integrating single-cell multi-omics data and inferring cross-omics regulatory relationships. scCross47 (https://github.com/mcgilldinglab/scCross). It is a deep generative model that unifies single-cell multi-omics data through modality-specific variational autoencoders and adversarial alignment, supporting integration of matched and unmatched data, cross-modal data generation, multi-omics simulation, and in silico perturbations. scCross47 (https://github.com/mcgilldinglab/scCross). It is a deep generative model that unifies single-cell multi-omics data through modality-specific variational autoencoders and adversarial alignment, supporting integration of matched and unmatched data, cross-modal data generation, multi-omics simulation, and in silico perturbations. scMM22 (https://github.com/kodaim1115/scMM). It is a mixture-of-experts deep generative model based on multi-modal variational autoencoders, designed for the integrated analysis of single-cell multi-omics data. scMM22 (https://github.com/kodaim1115/scMM). It is a mixture-of-experts deep generative model based on multi-modal variational autoencoders, designed for the integrated analysis of single-cell multi-omics data. MIDAS40 (https://github.com/labomics/midas). It is a deep generative modeling algorithm that disentangles biological states from technical noise to achieve unified clustering and knowledge transfer for mosaic single-cell multi-modal data with missing modalities and batch effects. MIDAS40 (https://github.com/labomics/midas). It is a deep generative modeling algorithm that disentangles biological states from technical noise to achieve unified clustering and knowledge transfer for mosaic single-cell multi-modal data with missing modalities and batch effects. multiDGD44 (https://github.com/Center-for-Health-Data-Science/multiDGD). It is a multi-modal deep generative decoder framework based on Gaussian mixture modeling, designed to learn disentangled low-dimensional representations of transcriptomic and epigenomic data for integrated analysis, clustering, and feature-level inference without requiring an encoder or prior feature selection. multiDGD44 (https://github.com/Center-for-Health-Data-Science/multiDGD). It is a multi-modal deep generative decoder framework based on Gaussian mixture modeling, designed to learn disentangled low-dimensional representations of transcriptomic and epigenomic data for integrated analysis, clustering, and feature-level inference without requiring an encoder or prior feature selection. scAI42 (https://github.com/sqjin/scAI). It is an unsupervised iterative learning method based on matrix factorization that integrates single-cell transcriptomic and epigenomic data through a cell aggregation strategy. scAI42 (https://github.com/sqjin/scAI). It is an unsupervised iterative learning method based on matrix factorization that integrates single-cell transcriptomic and epigenomic data through a cell aggregation strategy. MultiVI45 (https://scvi-tools.org/). It is a deep generative model that learns multi-modal embedding representations through joint analysis for integrating multi-modal and single-modality single-cell datasets and correcting technical variations such as batch effects. MultiVI45 (https://scvi-tools.org/). It is a deep generative model that learns multi-modal embedding representations through joint analysis for integrating multi-modal and single-modality single-cell datasets and correcting technical variations such as batch effects. scMVP32 (https://github.com/bm2-lab/scMVP). It is a multi-modal deep generative model based on VAE and GMM prior with transformer self-attention mechanism, designed to process single-cell joint profiling data that simultaneously measures gene expression and chromatin accessibility. scMVP32 (https://github.com/bm2-lab/scMVP). It is a multi-modal deep generative model based on VAE and GMM prior with transformer self-attention mechanism, designed to process single-cell joint profiling data that simultaneously measures gene expression and chromatin accessibility. MIRA46 (https://github.com/cistrome/MIRA). It is a probabilistic model combining cell-level topic modeling and gene-level RP modeling to analyze spatiotemporal dynamics of single-cell transcription and chromatin accessibility data. MIRA46 (https://github.com/cistrome/MIRA). It is a probabilistic model combining cell-level topic modeling and gene-level RP modeling to analyze spatiotemporal dynamics of single-cell transcription and chromatin accessibility data. Schema43 (https://github.com/rs239/schema). It is a multi-modal data analysis method based on metric learning and quadratic programming, designed to integrate heterogeneous single-cell data. Schema43 (https://github.com/rs239/schema). It is a multi-modal data analysis method based on metric learning and quadratic programming, designed to integrate heterogeneous single-cell data. scMoMaT50 (https://github.com/PeterZZQ/scMoMaT). It is a matrix-factorization-based single-cell data integration method specifically designed for mosaic integration scenarios, where input datasets may only partially overlap in features or cells. scMoMaT50 (https://github.com/PeterZZQ/scMoMaT). It is a matrix-factorization-based single-cell data integration method specifically designed for mosaic integration scenarios, where input datasets may only partially overlap in features or cells. MOJITOO35 (https://github.com/CostaLab/MOJITOO). It is a method based on canonical correlation analysis, designed for integrating multi-modal single-cell data with minimal parameter tuning requirements by learning a shared latent space and enabling the interpretation of feature relevance across modalities. MOJITOO35 (https://github.com/CostaLab/MOJITOO). It is a method based on canonical correlation analysis, designed for integrating multi-modal single-cell data with minimal parameter tuning requirements by learning a shared latent space and enabling the interpretation of feature relevance across modalities. SCOIT36 (https://github.com/deepomicslab/SCOIT) is a probabilistic tensor-decomposition framework for single-cell multi-omics integration. It models counts via Gaussian, Poisson or negative-binomial likelihoods to cope with sparsity and noise, and factorizes the data into joint cell, gene, and modality-specific latent embeddings. SCOIT36 (https://github.com/deepomicslab/SCOIT) is a probabilistic tensor-decomposition framework for single-cell multi-omics integration. It models counts via Gaussian, Poisson or negative-binomial likelihoods to cope with sparsity and noise, and factorizes the data into joint cell, gene, and modality-specific latent embeddings. scArches37 (https://github.com/theislab/scarches) is a deep learning framework based on transfer learning and parameter optimization for mapping single-cell query datasets to reference atlases. scArches37 (https://github.com/theislab/scarches) is a deep learning framework based on transfer learning and parameter optimization for mapping single-cell query datasets to reference atlases. scVAEIT38 (https://github.com/jaydu1/scVAEIT). It is a probabilistic variational autoencoder with masking strategies designed for single-cell multi-modal mosaic integration and transfer learning. scVAEIT38 (https://github.com/jaydu1/scVAEIT). It is a probabilistic variational autoencoder with masking strategies designed for single-cell multi-modal mosaic integration and transfer learning. CiteFuse39 (https://github.com/SydneyBioX/CiteFuse) is a computational framework specifically designed for CITE-seq data analysis. It employs similarity network fusion to integrate RNA transcriptome and cell-surface protein modalities. CiteFuse39 (https://github.com/SydneyBioX/CiteFuse) is a computational framework specifically designed for CITE-seq data analysis. It employs similarity network fusion to integrate RNA transcriptome and cell-surface protein modalities. Evaluation metrics We employ Normalized Mutual Information (NMI)87, Adjust Rand Index (ARI)88, Adjusted Mutual Information (AMI)89, and Accuracy (ACC)90 as evaluation metrics to evaluate the performance of the proposed model in scMulti-omic data with ground truth labels. They provide a reliable indication of the similarity between predicted and ground-truth labels. The clustering result with a higher value denotes better performance. NMI and AMI quantify the mutual information shared between two clustering results, ranging from 0 to 1. In particular, NMI focuses on normalizing mutual information to reflect the similarity between the clustering result and the true label, whereas AMI emphasizes adjusting mutual information to address random assignment and class imbalance issues. ARI measures the agreement between two clustering results, ranging from − 1 to 1; ACC compares the matching of the predicted clustering label with the true label to provide the agreement, ranging from 0 to 1. Given πe the predicted label and πt the ground truth label. They can be computed as follows: where ne, nt are the cluster numbers in πe and πt, respectively. nei is the number of samples in the cluster i of πe,ntj is the number of samples in cluster j of πt, and nij is the intersection sample size between clusters i and j,E(⋅) is the expected mutual information between clusters i and j. n is the number of cells in the scMulti-omics data, map(⋅) is a mapping function, usually implemented by the Hungarian algorithm91; δ(⋅) is the indicator function. To assess the clustering performance on unlabeled scMulti-omics datasets, three metrics are used, including Average Silhouette Width (ASW)92, Davies-Bouldin index (DB)92, and Calinski-Harabasz index (CH)93. Each offers unique insights into different aspects of cluster quality. Finally, biological conservation metrics and batch correction metrics are used to evaluate the ability of each model on multi-batch datasets, including biological conservation metrics: Isolated labels score94, KMeans NMI, KMeans ARI, Silhouette label, and Conditional Local Inverse Simpson’s Index (cLISI)95, batch correction metrics: Silhouette batch, Inverse Simpson’s Index of Integration (iLISI)95, KBET score51, and Graph connectivity (GC)37. Silhouette label and Silhouette batch scores were computed using the Average Silhouette Width (ASW) method, applied separately to clustering and batch labels. These metrics enable us to quantify the effectiveness of benchmarking the methods’ performance in reducing batch effect and the ability to retain relevant biological signals following data integration. Each evaluation metric is described in detail in Supplementary Note 14. Identifying marker genes, peaks, and ADTs in clusters To identify cluster-specific markers, we employ the ‘FindAllMarkers’ function from the Seurat package28, setting only.pos = TRUE to focus on positive markers, logfc.threshold = 0.25 to filter markers with a minimum log-fold change of 0.25, and min.pct = 0.25 to retain markers expressed in at least 25% of cells within any given cluster. After identifying markers, we rank genes, peaks, and ADTs based on statistical significance and effect size, selecting the top five markers per cluster. To visualize marker expression across clusters, we generate heatmaps using the ‘pheatmap’ function, effectively capturing expression patterns and facilitating biological interpretation. Motif analysis To identify transcription factor (TF) motifs within peak regions, we utilized the Signac package96. Motif analysis was conducted by first generating a motif matrix using the ‘CreateMotifMatrix’ function, followed by the creation of a motif object with ‘CreateMotifObject’. The ‘AddMotifs’ function was then applied to integrate motif data and perform an in-depth analysis. Using the BSgenome.Mmusculus.UCSC.mm10 reference genome, we computed Position Frequency Matrices (PFMs) based on the JASPAR2020 database. To identify specific TF motifs within target peak regions, we applied the ‘FindMotifs’ function, enabling precise motif discovery and characterization. Functional enrichment analysis A comprehensive Kyoto Encyclopedia of Genes and Genomes (KEGG) enrichment analysis was performed using the clusterProfiler package97 (v4.15.0.3). For this analysis, the top 500 most variable genes were selected, and the ‘enrichKEGG’ function was applied with organism = mmu, pvalueCutoff = 0.05, and qvalueCutoff = 0.05. In addition, Gene Ontology (GO) enrichment analysis was conducted to identify significant biological processes, cellular components, and molecular functions associated with gene sets. Genes were mapped to their corresponding GO terms using the org.Hs.eg.db database, and enrichment analysis was performed using the ‘enrichGO’ function with a p-value threshold of 0.05 to determine significantly enriched terms. The results were filtered and visualized using ‘dotplot’, ‘barplot’, and ‘cnetplot’ functions, providing a comprehensive insight into the biological significance of the identified gene sets. CellChat analysis To investigate intercellular communication networks, we employed the CellChat package76 (v1.6.1). A CellChat object was first created using the ‘createCellChat’ function, integrating gene expression matrices with cell annotation information. Significant signaling genes were extracted using the ‘subsetData’ function, focusing on those essential for intercellular communication. To identify key receptors and ligands within each cell population, we applied ‘identifyOverExpressedGenes’ and ‘identifyOverExpressedInteractions’. These interactions were further mapped onto a protein-protein interaction (PPI) network using the ‘projectData’ function, enabling the projection of ligand-receptor expression values onto the broader PPI framework. The ‘computeCommunProb’ function was then used to quantify communication probabilities for each ligand-receptor pair, assessing the likelihood of cell-cell signaling through specific interactions. Finally, we visualized the communication network, highlighting key pathways and intercellular interactions among distinct cell populations. Cell trajectory analysis Cell trajectory analysis was performed using the Monocle package98 (v1.3.7) to infer potential cellular differentiation trajectories. A Monocle object was first constructed using the ‘newCellDataSet’ function, providing the basis for downstream analysis. To reduce dimensionality, we applied the ‘reduceDimension’ function with DDRTree as the selected method, ensuring an optimized representation of the data. Cells were then ordered along the inferred trajectory using the ‘orderCells’ function, establishing sequential relationships based on gene expression patterns. Finally, we visualized the reconstructed cell trajectories using the ‘plot_cell_trajectory’ function, enabling the interpretation of dynamic cellular transitions. Gene set variation analysis (GSVA) To investigate the differential activity of gene sets across distinct cell subpopulations, we performed GSVA using the ‘gsva’ package99 (v1.52.3). The ‘AverageExpression’ function was first applied to compute the mean expression level of each gene within individual subpopulations. Subsequently, GSVA enrichment analysis was conducted using the ‘gsva’ function to evaluate variations in biological pathway activity. The results were visualized with the ‘pheatmap’ function, facilitating the interpretation of dynamic changes in biological processes across subpopulations. Gene set enrichment analysis (GSEA) To investigate the biological relevance of cell type-specific gene expression changes, we performed GSEA using the clusterProfiler package97 (v4.15.0.3). Differentially expressed genes were identified using the Seurat package28 (v4.2.0). Specifically, the ‘FindAllMarkers’ function was applied to identify marker genes for each cluster (or subgroup) in the RNA expression data. Genes significantly upregulated in a given cluster (adjusted p < 0.05, with a minimum log2 fold-change threshold) were considered marker genes of that cluster. These marker genes were then converted to their corresponding ENTREZIDs identifiers using the ‘bitr’ function. Genes were ranked in descending order based on their average log2 fold-change (avg_log2FC) to generate a pre-ranked gene list. GSEA was carried out with the ‘gseKEGG’ function (KEGG pathway database) using this ranked list, and enrichment plots were generated with ‘gseaplot2’. This analysis yielded key enriched pathways, providing insight into the molecular processes characterizing each cell population. Visualization of cell type-specific gene activity from chromatin accessibility To identify and visualize cell subpopulation-specific gene activity patterns based on chromatin accessibility, we applied the ‘plotMarkerHeatmap’ function from the ArchR package12 (v1.0.2). Gene activity scores were computed using the GeneScoreMatrix, which estimates accessibility around each gene’s transcription start site and its nearby regulatory elements. For visualization, these scores were Z-score normalized across clusters on a per-gene basis, by subtracting the mean and dividing by the standard deviation of each gene’s scores across all clusters. The resulting heatmap highlights cell-type-specific enrichment patterns, where each row represents a cell subpopulation, and each column represents a marker gene. Warmer colors reflect higher chromatin accessibility and potential transcriptional activity. RNA isolation and quantitative real-time PCR analyses We used the RNA Easy Fast Tissue/Cell Kits (Cat. #DP451, TIANGEN, China) to isolate RNA from kidney cancer cells and human kidney cancer tissues, respectively. The quality and concentration of extracted RNA was detected by a NanoPhotometer (Cat. #N60, Implen, Germany). We obtained cDNA according to the protocol of Hifair® AdvanceFast 1st Strand cDNA Synthesis Kit (Cat. #11149ES60, Yeasen Biotechnology, China). Real-time quantitative PCR (qRT-PCR) was performed using the Hieff® qPCR SYBR® Green Master Mix (Cat. #11201ES08, Yeasen Biotechnology, China) following the manufacturer’s instructions. The primers used for qPCR in this study are as follows: CSPP1 - Forward, 5’-GCCGCTGTAACCTCTTCGG-3’; CSPP1 - Reverse, 5’-GTCTTCGGCCAATCTGGCTTT-3’; LACTB2 - Forward, 5’-AAAACTCCCACGGAATCCTCA-3’; LACTB2 - Reverse, 5’-AATACCGTTGTTCCTTCCCCT-3’; LYPLA1 - Forward, 5’-CCTTTGCAGGTATCAGAAGTTCA-3’; LYPLA1 - Reverse, 5’-GCTGCCTGTTTAATCCCAGAT-3’; NCOA2 - Forward, 5’-TGGGGCCTATGATGCTTGAG-3’; NCOA2 - Reverse, 5’-GGTTTTTGACAAATTCCGTGTGG-3’; ADHFE1 - Forward, 5’-TGGACTTTCACCTTCTGGGAA-3’; ADHFE1 - Reverse, 5’-GGAGAGGTTCTTGTCTGTCATCA-3’. Quantitative functional validation of the role of HNF4A siRNA and transfection: siHNF4A compounds were all synthesized by GenePharma (Suzhou, China). The siRNA sequences were as follows: 5’-GAUCAGCACUCGAAGGUCATT-3’ (siHNF4A-1) and 5’-UCAUCUUCUUUGACCCAGATT-3’ (siHNF4A-2). Plated cells were transfected with a mixture of siRNA and Lipofectamine 3000 (Cat. #L3000015, Invitrogen) following the manufacturer’s instructions. CCK8 assay: Two thousand cells/well were seeded in 96-well plates. Following the indicated treatments, 10 μL CCK8 solution was added to the cell culture in each well at one-day intervals. After incubation for one hour, the absorbance values of the solutions were measured using a microplate reader. Colony-forming assay: Cells were plated in six-well plates (1000 cells/well) and cultured for 7–10 days until visible colonies formed. Colonies were fixed with 4% formaldehyde and then stained with 0.1% crystal violet for subsequent colony counting. Transwell migration assay: For migration assays, thirty thousand cells were seeded in an upper chamber (Corning) with serum-free mediums. After incubation for 24 h, the cells were fixed and stained with paraformaldehyde and crystal violet, respectively. Statistics & reproducibility This methodological study benchmarked scMAGCA on publicly available single-cell multi-omics datasets and validated it in kidney cancer cell lines and paired tumor and paracancerous tissues from seven patients. No statistical method was used to predetermine sample size. Sample sizes were determined by the availability of public datasets and biological materials. Data from prespecified development datasets, used solely for hyperparameter selection, were excluded from aggregated held-out benchmark summaries to avoid circularity; all other data were included in the analyses. The experiments were not randomized. The Investigators were not blinded to allocation during experiments and outcome assessment. Statistical analyses were two-sided unless otherwise stated, using t tests for Figs. 2a, b, d, 8g, and h, Wilcoxon rank-sum tests for differential expression, and one-sided hypergeometric tests for functional enrichment, with Benjamini-Hochberg false discovery rate correction where applicable. To support reproducibility, scMAGCA was run 10 independent times per dataset with different random seeds, all source code is publicly available, and replicate numbers for experimental assays are indicated in the corresponding figure legends. Reporting summary Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article. Supplementary information Supplementary Information Reporting Summary Transparent Peer Review file Source data Source Data Acknowledgements Figures 1a, b, d, f, h, i were created in BioRender. Zhou, Z. (2026) https://BioRender.com/t2ctudy. The authors also acknowledge support from the Hebei Province Key Laboratory of Big Data Calculation (Y.W., Z.Z., P.Z.). Author contributions X.L. conceived and supervised the project. Y.W. developed and implemented the algorithms under the guidance of X.L., Y.W. wrote the manuscript Y.W., Z.Z., and W.L. conducted the experiments. Y.W., X.L., and W.L. did the biological interpretation. Y.W., Z.Z., and Y.S. completed the figures and manuscript. Y.W., X.L., and Y.S. revised the manuscript. Y.C., K.-C.W., P.Z., and F.W. provided advice on method development. All authors approved the manuscript. Peer review Peer review information Nature Communications thanks Wei Chen and the other, anonymous, reviewer(s) for their contribution to the peer review of this work. A peer review file is available. Funding X.L. discloses support for the research and publication of this work from the National Natural Science Foundation of China [grant numbers 62472195 and 62076109], the Natural Science Foundation of Jilin Province [grant number 20260102302JC], and the Noncommunicable Chronic Diseases-National Science and Technology Major Project [grant numbers 2026ZD0554000 and 2026ZD0554005]. W.L. discloses support for the research of this work from the National Natural Science Foundation of China [grant number 82302920]. Y.W. discloses support for the research of this work from the Backbone Talent Program (Platform for Returned Overseas Scholars) [grant number A2025004] and the Langfang Science and Technology Plan Project [grant number 2025013056]. Y.S. discloses support for the research of this work from the Fundamental Research Funds for the Central Universities [grant number 135115035], Northeast Normal University. Z.Z., P.Z., Y.C., F.W., and K.W. declare no relevant funding. Data availability We compiled a diverse collection of publicly available single-cell multi-omics datasets spanning multiple species, tissues, and experimental platforms to enable comprehensive benchmarking. The full suite comprises 18 RNA+ADT datasets, 20 RNA+ATAC datasets, and one RNA+ADT+ATAC dataset. Sources and dataset ID are provided in Supplementary Table 20, which also summarizes for each dataset the name, data type, species, tissue, cell number, ADT/ATAC feature number, gene number, and label information. To ensure a strict separation between hyperparameter selection and headline benchmarking, we designate ten representative datasets as a development set for selecting default hyperparameters using label-free criteria: five RNA+ADT datasets (Dataset IDs 1–5) and five RNA+ATAC datasets (Dataset IDs 19–23). These development datasets are used only for label-free hyperparameter selection and are excluded from all aggregated benchmark statistics reported in the main text, namely average-rank and overall p-value summaries, while per-dataset results for the full dataset suite are still reported for transparency and reproducibility, with development vs. held-out benchmark datasets clearly annotated. The raw data analyzed in this study are publicly available from their respective original sources. Specifically, the 10X1kpbmc dataset is available at https://www.10xgenomics.com/datasets/1-k-pbm-cs-from-a-healthy-donor-gene-expression-and-cell-surface-protein-3-standard-3-0-0. The 10X5kpbmc dataset is available at https://support.10xgenomics.com/single-cellgene-expression/datasets/3.0.2/5k_pbmc_protein_v3. The 10X5kpbmcTotalSeq dataset is available at https://www.10xgenomics.com/datasets/5-k-peripheral-blood-mononuclear-cellspbm-cs-from-a-healthy-donor-with-cell-surface-proteins-next-gem-3-1-standard-3-1-0. The spector dataset is available at https://github.com/xianglin226/scMDC/tree/master/datasets. The zenodo6348128 dataset is available at https://zenodo.org/record/6348128. The 10Xmalt dataset is available at https://support.10xgenomics.com/single-cell-geneexpression/datasets/3.0.0/malt_10k_protein_v3. The GSE128639, GSE148127, GSE144744, GSE163120, GSE164378, GSE175702, GSE126074, GSE201402, and GSM4949911 datasets are available at GSE128639, [GSE148127], 32 [GSE144744], [GSE163120], [GSE164378], [GSE175702], [GSE126074], [GSE201402], and [GSM4949911], respectively. The GSE194122_s3d6 and GSE194122_s4d8 datasets are both available at [GSE194122]. The GSE150599_spleen_lymph_111 and GSE150599_spleen_lymph_206 datasets are both available at [GSE150599]. The GSE156478_Control, GSE156478_Stim, and GSE156478 datasets are all available at [GSE156478]. The GSE158013 dataset and the GSE158013 multiome dataset are both available at [GSE158013]. The human_ brain_3k dataset is available at https://www.10xgenomics.com/datasets/frozenhuman-healthy-brain-tissue-3-k-1-standard-2-0-0. The human_pbmc_3k dataset is available at https://www.10xgenomics.com/datasets/pbmc-from-a-healthy-donorgranulocytes-removed-through-cell-sorting-3-k-1-standard-2-0-0. The mouse_brain_5k dataset is available at https://www.10xgenomics.com/resources/datasets/freshembryonic-e-18-mouse-brain-5-k-1-standard-2-0-0. The human_brain_10x dataset is available at https://www.10xgenomics.com/resources/datasets/frozen-humanhealthy-brain-tissue-3-k-1-standard-2-0-0. The human_pbmc_10x dataset is available at https://www.10xgenomics.com/resources/datasets/pbmc-from-a-healthy-donorgranulocytes-removed-through-cell-sorting-3-k-1-standard-2-0-0. The E-MTAB-11264 dataset is available at https://www.ebi.ac.uk/biostudies/arrayexpress/studies/EMTAB-11264. The pbmc_granulocyte_sorted_10k dataset is available at https://www.10xgenomics.com/resources/datasets/pbmc-from-a-healthy-donor-granulocytesremoved-through-cell-sorting-10-k-1-standard-2-0-0. The pbmc_unsorted_10k dataset is available at https://www.10xgenomics.com/resources/datasets/pbmc-from-a-healthydonor-no-cell-sorting-10-k-1-standard-2-0-0. The pbmc_unsorted_3k dataset is available at https://www.10xgenomics.com/resources/datasets/pbmc-from-a-healthy-donorno-cell-sorting-3-k-1-standard-2-0-0. The pbmc_10X_public dataset is available at https://github.com/myylee/benchmark_sc_multiomic_integration. The 10X-Multiome-pbmc10k dataset is available at https://github.com/gao-lab/GLUE/tree/master/data/download/10x-Multiome-Pbmc10k. The Mouse_Kidney dataset is available at https://www.10xgenomics.com/datasets/mouse-kidneynuclei-isolated-with-chromium-nuclei-isolation-kit-saltyez-protocol-and-10x-complex-tissue-dp-ctsorted-and-ct-unsorted-1-standard. The Multiome_Mouse_Brain_Alzheimers dataset is available at https://www.10xgenomics.com/datasets/multiomic-integration-neuroscience-application-notesingle-cell-multiome-rna-atac-alzheimers-disease-mouse-model-brain-coronal-sections-from-onehemisphere-over-a-time-course-1-standard. Data supporting the results of this study can be found at 10.6084/m9.figshare.30899447(100). All other relevant data supporting the key findings of this study are available within the article and its Supplementary Information files. Source data are provided in this paper. Code availability The MIT-licensed source code of the scMAGCA software is released as a Python package at https://pypi.org/project/scMAGCAand is freely available on GitHub at https://github.com/zemingzhou30/scMAGCAand on Zenodo at https://zenodo.org/records/19234320101. To enhance the code availability and reproducibility, we have provided more detailed explanations in the GitHub source. We provided related demonstrations of how to use scMAGCA in various types of multi-omics datasets (RNA+ADT, RNA+ATAC, multi-batch dataset, and tri-omics dataset) at https://github.com/zemingzhou30/scMAGCA/tree/main/tutorial. In addition, https://github.com/zemingzhou30/scMAGCA/tree/main/figuresprovides code to reproduce the figures presented in this work, including the main benchmarks and the related biological analyses, facilitating independent verification and assessment. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. These authors contributed equally: Yunhe Wang, Zeming Zhou, Wei Liu. Supplementary information The online version contains supplementary material available at 10.1038/s41467-026-73055-7. References Associated Data Supplementary Materials Supplementary Information Reporting Summary Transparent Peer Review file Source Data Data Availability Statement We compiled a diverse collection of publicly available single-cell multi-omics datasets spanning multiple species, tissues, and experimental platforms to enable comprehensive benchmarking. The full suite comprises 18 RNA+ADT datasets, 20 RNA+ATAC datasets, and one RNA+ADT+ATAC dataset. Sources and dataset ID are provided in Supplementary Table 20, which also summarizes for each dataset the name, data type, species, tissue, cell number, ADT/ATAC feature number, gene number, and label information. To ensure a strict separation between hyperparameter selection and headline benchmarking, we designate ten representative datasets as a development set for selecting default hyperparameters using label-free criteria: five RNA+ADT datasets (Dataset IDs 1–5) and five RNA+ATAC datasets (Dataset IDs 19–23). These development datasets are used only for label-free hyperparameter selection and are excluded from all aggregated benchmark statistics reported in the main text, namely average-rank and overall p-value summaries, while per-dataset results for the full dataset suite are still reported for transparency and reproducibility, with development vs. held-out benchmark datasets clearly annotated. The raw data analyzed in this study are publicly available from their respective original sources. Specifically, the 10X1kpbmc dataset is available at https://www.10xgenomics.com/datasets/1-k-pbm-cs-from-a-healthy-donor-gene-expression-and-cell-surface-protein-3-standard-3-0-0. The 10X5kpbmc dataset is available at https://support.10xgenomics.com/single-cellgene-expression/datasets/3.0.2/5k_pbmc_protein_v3. The 10X5kpbmcTotalSeq dataset is available at https://www.10xgenomics.com/datasets/5-k-peripheral-blood-mononuclear-cellspbm-cs-from-a-healthy-donor-with-cell-surface-proteins-next-gem-3-1-standard-3-1-0. The spector dataset is available at https://github.com/xianglin226/scMDC/tree/master/datasets. The zenodo6348128 dataset is available at https://zenodo.org/record/6348128. The 10Xmalt dataset is available at https://support.10xgenomics.com/single-cell-geneexpression/datasets/3.0.0/malt_10k_protein_v3. The GSE128639, GSE148127, GSE144744, GSE163120, GSE164378, GSE175702, GSE126074, GSE201402, and GSM4949911 datasets are available at GSE128639, [GSE148127], 32 [GSE144744], [GSE163120], [GSE164378], [GSE175702], [GSE126074], [GSE201402], and [GSM4949911], respectively. The GSE194122_s3d6 and GSE194122_s4d8 datasets are both available at [GSE194122]. The GSE150599_spleen_lymph_111 and GSE150599_spleen_lymph_206 datasets are both available at [GSE150599]. The GSE156478_Control, GSE156478_Stim, and GSE156478 datasets are all available at [GSE156478]. The GSE158013 dataset and the GSE158013 multiome dataset are both available at [GSE158013]. The human_ brain_3k dataset is available at https://www.10xgenomics.com/datasets/frozenhuman-healthy-brain-tissue-3-k-1-standard-2-0-0. The human_pbmc_3k dataset is available at https://www.10xgenomics.com/datasets/pbmc-from-a-healthy-donorgranulocytes-removed-through-cell-sorting-3-k-1-standard-2-0-0. The mouse_brain_5k dataset is available at https://www.10xgenomics.com/resources/datasets/freshembryonic-e-18-mouse-brain-5-k-1-standard-2-0-0. The human_brain_10x dataset is available at https://www.10xgenomics.com/resources/datasets/frozen-humanhealthy-brain-tissue-3-k-1-standard-2-0-0. The human_pbmc_10x dataset is available at https://www.10xgenomics.com/resources/datasets/pbmc-from-a-healthy-donorgranulocytes-removed-through-cell-sorting-3-k-1-standard-2-0-0. The E-MTAB-11264 dataset is available at https://www.ebi.ac.uk/biostudies/arrayexpress/studies/EMTAB-11264. The pbmc_granulocyte_sorted_10k dataset is available at https://www.10xgenomics.com/resources/datasets/pbmc-from-a-healthy-donor-granulocytesremoved-through-cell-sorting-10-k-1-standard-2-0-0. The pbmc_unsorted_10k dataset is available at https://www.10xgenomics.com/resources/datasets/pbmc-from-a-healthydonor-no-cell-sorting-10-k-1-standard-2-0-0. The pbmc_unsorted_3k dataset is available at https://www.10xgenomics.com/resources/datasets/pbmc-from-a-healthy-donorno-cell-sorting-3-k-1-standard-2-0-0. The pbmc_10X_public dataset is available at https://github.com/myylee/benchmark_sc_multiomic_integration. The 10X-Multiome-pbmc10k dataset is available at https://github.com/gao-lab/GLUE/tree/master/data/download/10x-Multiome-Pbmc10k. The Mouse_Kidney dataset is available at https://www.10xgenomics.com/datasets/mouse-kidneynuclei-isolated-with-chromium-nuclei-isolation-kit-saltyez-protocol-and-10x-complex-tissue-dp-ctsorted-and-ct-unsorted-1-standard. The Multiome_Mouse_Brain_Alzheimers dataset is available at https://www.10xgenomics.com/datasets/multiomic-integration-neuroscience-application-notesingle-cell-multiome-rna-atac-alzheimers-disease-mouse-model-brain-coronal-sections-from-onehemisphere-over-a-time-course-1-standard. Data supporting the results of this study can be found at 10.6084/m9.figshare.30899447(100). All other relevant data supporting the key findings of this study are available within the article and its Supplementary Information files. Source data are provided in this paper. The MIT-licensed source code of the scMAGCA software is released as a Python package at https://pypi.org/project/scMAGCAand is freely available on GitHub at https://github.com/zemingzhou30/scMAGCAand on Zenodo at https://zenodo.org/records/19234320101. To enhance the code availability and reproducibility, we have provided more detailed explanations in the GitHub source. We provided related demonstrations of how to use scMAGCA in various types of multi-omics datasets (RNA+ADT, RNA+ATAC, multi-batch dataset, and tri-omics dataset) at https://github.com/zemingzhou30/scMAGCA/tree/main/tutorial. In addition, https://github.com/zemingzhou30/scMAGCA/tree/main/figuresprovides code to reproduce the figures presented in this work, including the main benchmarks and the related biological analyses, facilitating independent verification and assessment.