LONMF: a non-negative matrix factorization model based on graph Laplacian and optimal transmission for paired single-cell multi-omics data integration Authors: Mengdi Nan, Qing Ren, Yuhan Fu, Xiang Chen, Guanpeng Qi, Liugen Wang, Jie Gao Published: 2025 ABSTRACT AbstractThe rapid development of single-cell sequencing technologies has provided a robust technical support for the efficient resolution of multiple levels of molecular information from a single-cell population. However, the data produced by these technologies often contain a lot of noise and differences in characteristics that make it difficult to integrate and analyze single-cell multi-omics data. In this study, there is a growing demand for methods to integrate single-cell multi-omics data, which is expected to enhance the ability to reveal cellular heterogeneity and provide new biological perspectives for a deeper understanding of cellular phenotypes by jointly analyzing multi-omics data. We propose LONMF, a non-negative matrix factorization algorithm combining graph Laplacian and optimal transmission to enhance clustering performance and interpretability. We apply LONMF to visualize and cluster multi-pair single-cell multi-omics data, including 10X-multi-group, CITE-seq, and TEA-multi-group seq, to facilitate marker characterization and gene ontology enrichment analysis and to provide rich biological insights for downstream analyses. Our comprehensive benchmarking demonstrates that LONMF exhibits comparable performance compared with the current state-of-the-art in cell clustering and outperforms other methods in terms of biological interpretability.Supplementary InformationThe online version contains supplementary material available at 10.1186/s12859-025-06301-2.Keywords: Single-cell multi-omics, Paired data, Non-negative matrix decomposition, Data integration, Optimal transmission, Graph laplacian Introduction With the rapid development of single-cell multi-omics technologies, we can simultaneously capture and integrate multiple molecular features of the genome, transcriptome, and proteome at the single-cell level [1–4]. These technologies have significantly improved our ability to resolve cellular states while deepening our understanding of the complexity of biological systems [5–9]. In particular, acquiring paired multi-omics data with multiple molecular signatures from the same cell opens up new possibilities for integrative analyses. These paired multi-omics data have helped us to reveal the dynamics of cellular functions and states and have facilitated in-depth exploration of cellular heterogeneity, gene regulation, and developmental mechanisms [10–13]. Due to the high dimensionality and complexity of single-cell multi-omics data, there is an urgent need to develop analytical methods capable of integrating these paired data to accurately capture the synergistic relationships between molecular levels. Based on this, integration methods developed in recent years can accurately capture associations between multi-omics levels, such as regulatory networks of gene expression or protein functions concerning genomic changes [14, 15], while significantly improving the accuracy of analyses and reducing the noise and bias of traditional methods. To this end, many advanced methods for integrating unpaired and paired single-cell multi-omics data have been developed [16–21]. For unpaired single-cell multi-omics data integration methods, such as LIGER [16], integral non-negative matrix factorization (iNMF) is employed to integrate data. Due to the non-convex nature of the problem, results may be sensitive to initialization. This method identifies shared cell types across individuals, species, and modalities (e.g., gene expression, epigenetics, spatial transcriptomics) while capturing dataset-specific features, enabling unified analysis of heterogeneous single-cell data. GLUE [17] is a modular integration framework that handles unpaired single-cell multi-omics data and infers cross-omics regulatory relationships. By explicitly modeling regulatory networks, GLUE establishes biologically intuitive connections across different omics feature spaces. eMCI [18] is an interpretable deep neural network-based multimodal integration model that fuses scRNA-seq and spatial transcriptomics data. It excels in cell type classification and deconvolution tasks, identifying key spatial features through attribution analysis and revealing cell type-specific and cell communication spatiotemporal patterns. For paired single-cell multi-omics data integration methods, such as MOFA + [19], which reconstructs a low-dimensional data representation using multi-omics factor analysis and efficient variational inference, while supporting flexible sparsity constraints for joint modeling across multiple sample groups and modalities. Cobolt [20] is a multimodal variational autoencoder (MVAE) based approach to generate modality-independent unified representations for the combined analysis of paired multi-omics data. Furthermore, Mowgli [21] is an approach that effectively integrates paired multi-omics data by combining non-negative matrix decomposition with optimal transmission strategies, thereby enhancing the clustering performance and interpretability of the results. These innovative approaches for paired multi-omics data can more efficiently mine inter-cell shared and complementary information than traditional unpaired data integration methods, thereby greatly advancing the theoretical and practical progress of single-cell multi-omics data integration. Non-negative matrix factorization (NMF) [22] is commonly used in single-cell multi-omics data integration to reduce high-dimensional data into a low-dimensional space, enabling the extraction of biologically interpretable features. Various extended NMF models have also been developed. For example, JNMF [23] projects heterogeneous data into a unified coordinate system, facilitating the construction of multidimensional modules; sNMF [24] incorporates semi-orthogonal constraints to enhance data representation and improve the interpretability of clustering results; and JSNMF [25] integrates transcriptomic and epigenomic data, using consensus map fusion to address the discrepancy between feature scales and noise levels. However, classical NMF is inherently limited by its linearity assumptions, which limits its ability to model complex nonlinear relationships and higher-order interactions in multi-omics data. The Laplace matrix is essential for the integration and analysis of single-cell multi-omics data. By constructing inter-cell graph structures, it captures similarities between samples, thereby revealing the data’s local geometric features and global topological relationships. Taking paired multi-omics analysis as an example, this method not only identifies shared features among cells but also preserves the modality specificity of the data. Particularly suited for high-dimensional, high-noise datasets, it constructs low-dimensional embeddings that reflect cellular heterogeneity, significantly enhancing the accuracy and biological interpretability of integrated analyses. In this paper, we present a new non-negative matrix factorization algorithm, LONMF, based on graph Laplacian and optimal transmission. This algorithm is designed for integrative analysis of paired single-cell multi-omics data. The method not only more efficiently captures the similarity between single-cell histological modalities to construct complex relationships between modalities but also reveals the high-dimensional geometric structure of each histology in the original data. The superior clustering performance of the method, as demonstrated by its effectiveness in analyzing datasets such as PBMC 10X, the Liu dataset, and CITE-seq, indicates its superiority over other state-of-the-art methods. The examination of the three modal datasets by LONMF not only accurately distinguishes the inter-specificity of immune cells but also yields profound biological insights into the classification of immune cell subpopulations. Materials and methods Data preprocessing The data preparation stage of this study is conducted using the Python packages Scanpy [26] and Muon [27]. The LONMF preparation of single-cell RNA sequencing gene expression profiles proceeds as follows: initially, quality control, gene filtering, and cell filtering based on bespoke criteria are performed. Subsequently, the gene expression matrix is normalized and log-transformed, and the genes exhibiting the greatest variability are identified. The preparation of scATAC-seq expression profiles mirrors that of scRNA-seq, with the the following steps: initially, quality control of cells and peak and cell filtering are conducted according to bespoke criteria. Subsequently, the peak expression matrix is normalized and log-transformed, and peaks exhibiting high variability are identified. The preparation of scADT-seq expression profiles involves solely normalizing the data without implementing quality control, protein, or cell filtering. Due to the biological and technical nature of the datasets, the first 1500 highly variable genes and 1500 peaks are selected for the Liu dataset, the first 2500 highly variable genes and 5000 peaks are selected for the PBMC 10X(RNA + ATAC) dataset, and for the TEA dataset, 2500 highly variable genes and 15,000 peaks are selected. Similarly, for the OPCITE and PBMC 10X(RNA + ADT) datasets, the first 2500 highly variable genes and 15,000 peaks are assigned. LONMF framework Nonnegative matrix factorization For a given non-negative data matrix , the application of the classical non-negative matrix factorization model [21] aims to find and that approximate the matrix of , where is the potential dimension. The objective function of the nonnegative matrix factorization (NMF) model is shown below: where means that the matrices are non-negative, and is the Frobenius paradigm, which is used to measure the differences between matrices. In the work of integrating paired data based on NMF framework, Mowgli model [21] combines non-negative matrix factorization and entropy-regularized optimal transport (OT) to enhance clustering and interpretability. It decomposes each matrix into a modality-specific dictionary matrix and a shared embedding matrix , with simplex constraints ensuring the column sums of and equal 1 to maintain valid entropy values. The complete loss function for Mowgli combines the reconstruction error, the entropy regularisation term, and the simplex constraint in the form: where the optimal transmission for entropy regularisation is defined as: where the entropy term E(P) is expressed as:. The introduction of entropy regularisation not only reduces the complexity of OT calculations but also enhances numerical stability. The ground cost momentsare calculated by the cosine distance between features: where denotes the inner product of the feature vectors and , and and are the second-paradigm numbers of the vector and respectively. This definition effectively reflects the similarity between features and provides an important basis for the calculation of the optimal transmission distance, thus providing a more flexible and efficient tool to support the integration and analysis of multi-omics data. Graph laplace construction In single-cell data analysis, the objective is to guarantee that two cells situated nearby in the original data space are also positioned nearby in the low-dimensional potential space. To achieve this, Laplacian plots are employed to capture the high-dimensional geometric structure of each data set. The Laplace matrix, denoted by , is defined as , where is the degree matrix and is the adjacency matrix. The adjacency matrix calculates the similarity between units using an exponential similarity kernel function, whose formula is as follows: where the parameter is empirically set to 0.5 and is used to eliminate the scale difference problem. The symbol denotes the square of the Euclidean distance between cells and . The symbol denotes the average of the squares of the Euclidean distances between cell and its 20 nearest neighbors, and similarly, denotes the average of the squares of the Euclidean distances between cell and its 20 nearest neighbors. The Laplacian matrix plays a crucial role in integrating and analyzing single-cell multi-omics data. Its construction process involves treating cells as nodes in a graph and establishing weighted edges based on molecular feature similarity, thereby generating an adjacency matrix and computing the Laplacian matrix. With its unique spectral properties, it can simultaneously characterize both the local geometric structure and global topological relationships of the data, demonstrating outstanding performance in tasks such as dimensionality reduction, clustering, and feature selection. In multi-omics integration, the Laplacian matrix captures common structural patterns across different modalities while preserving their distinct characteristics, thereby providing robust support for unified modeling of complex heterogeneous data. LONMF model The LONMF model combines optimal transport with graph Laplacian, drawing inspiration from the Mowgli model [21], to enhance clustering performance while retaining the structure of high-dimensional data. For the normalized single-cell data matrix (where is the number of features and is the number of cells), the model firstly extracts the same cells from it and decomposes each data matrix into a modality-specific dictionary matrix and the product of the embedding matrix shared across modalities, where denotes the dimensionality of the latent space and is set to a value much smaller than the number of features. LONMF’s objective function takes the following form: where is the modal dictionary specific to the data matrix of the single-cell dataset, is the cell embeddings shared across modalities, and is the Laplace diagram, which captures the high-dimensional geometric structure in the original data space. The is the weight parameter in the p-th Laplace diagram and is adaptively learned from the data. Regularization is the standardized term for LONMF, which contains regularization for both cell embedding and feature embedding. The framework diagram of the LONMF is shown in Fig. 1. Schematic overview of LONMF. By preprocessing the single-cell multi-omics data, the corresponding data matrices are obtained.LONMF takes the histological data matrices as inputs, and obtains the feature matrices and cell matrices by constructing the graph Laplacian and optimal transmission algorithms. The loss functions of the model include optimal transmission loss and graph loss. The latent representation is used for subsequent visualization and clustering By combining the entropy term and reconstruction with LONMF’s model, the loss function is In the objective function (8) of LONMF, the first term, represents the OT-based NMF decomposition model, which is used to calculate the NMF loss for the single-cell dataset. The second and third terms in the loss function (8) represent the entropy functions of and , respectively. The fourth term of the objective function is designed to enhance the performance of the model, and the Laplace map captures the high-dimensional geometric structure in the original data space of the th data modality. As in the study of Rolet et al. [28], the optimization problem of and is transformed into a smooth minimization problem on the newly introduced dyadic variable when minimizing Eq. (8). This equivalent transformation permits the problem to be solved by standard optimization methods, whereby the L-BFGS algorithm is employed. Optimize to solve the following smooth minimum problem: After that, update the original variables as follows: where the softmax function is defined as Optimize to solve the following smooth minimum problem: After that, update the original variables as follows: where and are Legendre dyadic functions of and , respectively. The pseudocode for the LONMF algorithm is as follows: LONMF algorithm pseudocode Parameter selection In order to make these parameters more comparable across histologies and datasets, we define: In the experiment based on Mowgli [21], we configured the parameters as , , and . Meanwhile, the settings of and are determined based on their best performances in the dataset and in the assessment metrics (e.g., silhouette scores, ARI, and purity scores) (see Supplementary Figs. S1, S2, and S3 of Supplementary Materials 1). Summary information on symbols used in the Methods section is provided in Supplementary Table S2 of Supplementary Materials 1. Evaluation metrics i)Silhouette Score. For each sample, the Silhouette score is defined as Silhouette Score. For each sample, the Silhouette score is defined as where is the average distance of the sample to other samples in the same cluster and is the average distance of the sample to the most recently clustered sample. It is a metric used to evaluate the clustering results and helps to measure how closely the data points belong to the clusters to which they belong and how well the different clusters are separated from each other. It is commonly used for evaluating the effectiveness of unsupervised learning algorithms such as K-means and hierarchical clustering. Its value ranges from − 1 to 1, with higher values indicating better clustering results. ii) KNN score. KNN purity score measures the average proportion of nearest neighbors of samples shared with the sample clustering annotation. Thus, it varies between 0 and 1. KNN score. KNN purity score measures the average proportion of nearest neighbors of samples shared with the sample clustering annotation. Thus, it varies between 0 and 1. iii) Adjusted Rand Index. The Rand index defines the similarity between real annotations and experimental clustering. ARI is defined as Adjusted Rand Index. The Rand index defines the similarity between real annotations and experimental clustering. ARI is defined as where Rand index is represented by . The ARI varies between − 1 and 1, with 0 indicating a random clustering. Comparative methods MOFA + [19] extends MOFA for large-scale multi-omics analysis with GPU acceleration and sparsity priors but may lose critical information due to its factor screening parameters. Therefore, we decided to keep 5 factors for Liu and the Liu dataset, and 15 factors for the other datasets. Cobolt [20] integrates multimodal data using a Bayesian model and variational autoencoder. We trained its Python implementation with a 0.001 learning rate for 200 epochs. For Liu and Liu’s dataset, we use 5 latent dimensions, while for the other datasets, we use 30 latent dimensions. NMF [29] which we implemented using the TorchNMF package. For Liu and related datasets, we chose 5 latent dimensions, while 30 latent dimensions were used for the other datasets. Mowgli [21] integrates paired multi-omics data using non-negative matrix decomposition and optimal transport. We use 5 latent dimensions for Liu and Liu’s dataset and 50 latent dimensions for the other datasets. Results LONMF provides superior clustering performance for datasets while enhancing data visualization We benchmark LONMF, MOFA + , Integrated NMF, Cobolt, and Mowgli based on their embedding and clustering performance on these two single-cell multi-omics datasets (RNA + ADT) (Fig. 2a, c). We selected these two datasets across different sequencing technologies, tissues, and sizes: (i) PBMC 10X (RNA + ADT) is a 10X multi-omics human PBMC dataset from 10X Genomics; (ii) OP CITE is a human bone marrow dataset from Open Problems [30] CITE-seq. Supplementary Table S1 lists the patterns, number of cells, and cell types for each dataset used in this paper. For more information on data preprocessing, see Methods. Meanwhile, as shown in Fig. 2a, b, LONMF can process large single-cell multi-omics datasets and provide embedding and clustering performance with state-of-the-art clustering to investigate the profiles of protein modifications and gene expression in human cells. Based on silhouette scores, LONMF outperforms other methods in PBMC 10X (RNA + ADT) and OP CITE. Regarding ARI across resolutions, LONMF outperforms almost all other methods in the OP CITE; Integrated NMF performs best overall on the PBMC 10X dataset. Regarding purity scores, LONMF outperforms the different methods on the OP CITE and PBMC 10X datasets. Overall, LONMF achieves the highest silhouette scores and purity scores with Analysis of the OPCITE and PBMC10X(RNA + ADT) dataset. a Comparison of clustering results, evaluated by silhouette scores, ARI and purity scores in OPCITE dataset. b Comparison of clustering results, evaluated by silhouette scores, ARI and purity scores in PBMC10X(RNA + ADT) dataset. c UMAP visualization of real cell types in OPCITE dataset and PBMC10X(RNA + ADT) dataset good clustering performance (Fig. 2a, b). The UMAP plot in Fig. 2c gives a qualitative intuition of the performance of each method. In OP CITE, only integrated NMF, Cobolt, Mowgli, and LONMF correctly separate subpopulations of B cells. In PBMC 10X (RNA + ADT), LONMF and Mowgli separated better than the other methods. LONMF enables robust visualization and deep biological insights into cell subpopulations On the PBMC 10X dataset (RNA + ATAC), we compare the clustering performance of LONMF with MOFA + , integrated NMF, Cobolt, and Mowgli. Based on silhouette scores, LONMF outperforms the other methods on this dataset. In terms of Adjusted Rand Index (ARI) across resolutions, LONMF performs best, while in terms of purity scores, MOFA + shows a slight advantage over LONMF, and both are superior to other comparative methods. Overall, LONMF achieves the highest silhouette scores and a high purity score with good clustering performance. (Fig. 3a). In addition, the UMAP plot in Fig. 3b provides a qualitative comparison of the methods on the PBMC 10X (RNA + ATAC) dataset. As can be seen from the plots, the LONMF and Mowgli methods cluster cells of the same cell type more tightly, outperforming the other methods. Analysis of the PBMC10X(RNA + ATAC) dataset. a Comparison of clustering results, evaluated by silhouette scores, ARI and purity scores in BMC10X(RNA + ATAC) dataset. b UMAP visualization of real cell types in PBMC10X(RNA + ATAC) dataset. c The violin plot illustrates the expression levels of various cell types in factor 8 and factor 10. d A gene ranking plot for factor 8 and factor 10 In further analysis, we construct a heatmap of factors versus cell types (see Supplementary Fig. S4 of Supplementary Materials 1), which clearly illustrates the specific expression patterns of each factor across different cells. Based on this, we focus particularly on Factor 8 and Factor 10. Violin plots show that CD56 bright NK cells have higher values in factor 8, while memory B cells have higher values in factor 10 (Fig. 3c). Based on the gene scores defined by scAI [31], we plot the gene rankings (Fig. 3d). Specifically, for factor 8, we divide the expression value of each gene by the sum of the expression values of all genes in the factor, highlighting the 10 marker genes for CD56 bright NK cells in CellMarker 2.0 [32] (KLRF1, TRDC, NCAM1, SPON2, NCR1, FGFBP2, TBX21, CCL4, CD160, CD38) that scored high in factor 8. Similarly, for factor 10, we use the same calculation method to highlight 10 marker genes for memory B cells (ATM2, CD79A, FCRLA, MS4A1, TNFRSF13C, BANK1, CD79B, VPREB3, IGHF1, PAX-5), and the gene scores for these marker genes are also generally higher in factor 10. For factor 8, we rank genes based on expression levels and select the top 100 genes as specific genes. We then perform GO biological process (BP) enrichment analysis using the ClusterProfiler R package. The analysis employs the Benjamini–Hochberg method to adjust the corrected p-values from Fisher’s exact test, with results expressed as log-10 corrected p-values. The screening threshold is set at 0.001, identifying 96 significantly enriched biological processes (Supplementary Materials 2, Table 1). Results indicate that Factor 8 is significantly enriched in biological processes related to leukocyte-mediated immunity and cell surface receptor signaling involved in regulating the immune response. This finding aligns with the characteristics of the CD56-high natural killer (NK) cell subset. CD56-high NK cells regulate the functions of other immune cells by secreting immune factors such as IFN-γ and TNF-α [33]. Their cell surface receptors are regulated through signaling pathways, including PI3K-AKT and JAK-STAT, which play crucial roles in chronic inflammation, viral infections, and tumor immunity [34]. Consequently, they have become a key focus in immunotherapy research. Table 1GO BP enrichment analysis of Factor 8 of the PBMC 10X (RNA + ATAC) dataset (top 10 biological processes)IDID descriptionLog10(p.adjust)GO:0002443leukocyte mediated immunity− 9.04GO:0002768immune response-regulating cell surface receptor signaling pathway− 9.04GO:0001906cell killing− 9.04GO:0001909leukocyte mediated cytotoxicity− 8.56GO:0046631alpha–beta T cell activation− 8.56 GO BP enrichment analysis of Factor 8 of the PBMC 10X (RNA + ATAC) dataset (top 10 biological processes) For Factor 10, 48 significant biological processes are screened using the same method (Supplementary Materials3, Table 2). Results indicate that Factor 10 is significantly enriched in biological processes related to leukocyte-mediated immunity and cell surface receptor signaling involved in immune response regulation. Memory B cells constitute a crucial B cell subset within the adaptive immune system, primarily functioning to rapidly recognize and respond to previously encountered antigens. During secondary immune responses, memory B cells recognize antigens via high-affinity B-cell receptors (BCRs) and rapidly differentiate into antibody-secreting cells. Their function is closely linked to antigen receptor signaling pathways and BCR signaling pathways. BCR-mediated signaling activates downstream molecules (such as SYK and BTK), promoting cell proliferation, antibody secretion, and immune responses [35]. The BCRs of memory B cells possess higher affinity than those of naive B cells, enabling faster and more efficient responses during secondary immune responses. Table 2GO BP enrichment analysis of Factor 10 of the PBMC 10X (RNA + ATAC) dataset (top 10 biological processes)IDDescriptionLog10(p.adjust)GO:0050851antigen receptor-mediated signaling pathway− 15.60GO:0050853B cell receptor signaling pathway− 13.57GO:0002429immune response-activating cell surface receptor signaling pathway− 13.57GO:0002768immune response-regulating cell surface receptor signaling pathway− 12.91GO:0042113B cell activation− 12.62 GO BP enrichment analysis of Factor 10 of the PBMC 10X (RNA + ATAC) dataset (top 10 biological processes) In addition, we use the GREAT tool (http://great.stanford.edu/public/html/)to perform an online GO BP enrichment analysis of the top 1000 ranked loci in each factor. The filtering criteria are Binom raw P-value < 0.001 and Binom FDR Q-value < 0.001. The analysis shows that Factor 8 and Factor 10 are significantly enriched in several immune system-related biological processes, including immune response (Factor 8: ; Factor 10: ), regulation of immune system process (Factor 8: ; Factor 10: ) and positive regulation of immune system process (Factor 8: ; Factor 10: ). These enrichment results suggest that factor 8 and factor 10 may play critical functions in specific cell types, providing important functional insights for further studies. In a variety of data environments, LONMF performs well in cell embedding and clustering tasks We apply LONMF to the Liu dataset to evaluate its embedding and clustering performance, while several data transformations are performed for different distribution scenarios of the three groups of cells. To more realistically simulate the distribution scenarios of the three groups of cells, we apply different transformations to a simple dataset consisting of scCAT-seq. The scCAT-seq dataset provides a combined analysis of scRNA and scATAC for the HCT116, HeLa-S3, and K562 cell lines. Unlike simulated data, this approach avoids making assumptions about the data distribution. Generating simulated data according to a Gaussian distribution, for example, will tend to favor methods that approximate single-cell data with a similar distribution. We design three scenarios to model realistic challenges in multi-omics integration: (i) RNA mixing and mixing of the two scenarios, including two types of cell populations, one of which is mixed in scRNA but completely segregated in scATAC, and the other of which is mixed in scRNA but well segregated in scATAC, or mixed in scATAC but well segregated in scRNA; (ii) rare cell populations, i.e., cell populations that are significantly less numerous than other cell populations; (iii) 82\% sparse, 90\% sparse, and 96\% sparse. Scenario (i) is used to assess the ability of different methods to handle the complementarity of multi-omics data; scenario (ii) is used to test the effectiveness of the methods in recovering rare cell populations; and scenario (iii) is used to test the robustness of the methods to cope with dropout noise while preserving the true pattern of loss in single-cell data. Specific details on dataset generation can be found in the Methods section. We then benchmarked LONMF, Mowgli, MOFA + , NMF, and Cobolt (Fig. 4a, b, e). We also compare the effectiveness of different methods for UMAP visualization on the Liu dataset (Fig. 4c, d, f). Analysis of the Liu dataset. a Comparison of clustering results assessed by silhouette score, ARI and purity score in Scenario 1. b Comparison of clustering results assessed by silhouette score, ARI and purity score in Scenario 2. c UMAP visualization of real cell types in Scenario 1. d UMAP visualization of real cell types in Scenario 2. e Comparison of clustering results assessed by silhouette score, ARI and purity score in Scenario 3. f UMAP visualization of real cell types in Scenario 3 LONMF enhances biological interpretability and identifies key immune cell subpopulations in TEA-seq data through cell-type-specific factors We benchmark LONMF on the TEA-seq dataset of human PBMCs to evaluate its biointerpretability. The TEA-seq dataset is a single-cell, multi-omics dataset that contains information on gene expression, chromatin accessibility, and tagged proteins and provides an opportunity for method testing to fully account for the complementarity of molecular regulatory hierarchies. We independently apply MOFA + , Mowgli, and LONMF to the three histological integration analyses of the TEA-seq data and cluster the embedded results generated by each of the three methods. Since cellular annotations are not provided in the dataset, we annotate the clustered results with cell types in combination with gene and protein markers (Fig. 5a, b). This process successfully identifies major immune cell types, including CD4T cells, CD8T cells, B cells, natural killer (NK) cells, MAIT T cells, monocytes, and erythroid cells. Notably, the consistency of the cell type annotation results of MOFA + , Mowgli, and LONMF is as high as 98%, which is in high agreement with that of Azimuth’s RNA-based annotation results (see Fig. 5c), suggesting that each of these methods can accurately recover the expected cell types by clustering their embedded data. In addition, we quantify the performance of the LONMF factors using a specificity score based on the Mowgli definition, and the results are shown in bold in Fig. 5d. The analysis shows that LONMF performs well in establishing one-to-one associations between the factors and cell types, and the specificity scores stood out significantly, especially in NK cells, CD8T cells, and CD4T cells. The high selectivity of the LONMF factors efficiently captures the specific biological features of these cell types. Meanwhile, the specificity analysis of LONMF factors is robust to hyperparameter selection (see Supplementary Fig. S5 of Supplementary Materials 1), further validating its reliability in biological interpretability. Analysis of the TEA dataset. a Gene and protein markers for LONMF annotation of TEA-seq data. b UMAP visualization of LONMF embedding c Comparison of coarse annotations of TEA-seq data obtained through different methods. d The average weights within and outside a cell type are visualized for each factor in LONMF Finally, we focus on the biological relevance of the factors identified by LONMF in Human PBMC TEA-seq data. Based on Mowgli [21], LONMF is able to identify multiple factors associated with subpopulations of cell types, e.g., distinguishing between memory B cells and naive B cells, while capturing specific features of different subpopulations of CD8 T cells, monocytes, dendritic cells, and NK cells. The association of these factors with specific immune subpopulations is determined based on top genes and proteins in effector memory CD8 T cells, naive B cells, memory B cells, and CD56dim NK cells. For other cell populations, we determine associations by analyzing correlations between factor weights and weights of known protein markers. Specifically, we quantify LONMF factors associated with specific immune subpopulations in the TEA-seq PBMC dataset, including CD56 bright NK cells, naive CD4 T cells, plasmacytoid dendritic cells, and conventional dendritic cells. By comprehensively evaluating the ranking of top protein markers among factors, Pearson correlations between factors and protein markers, and scatter plots, we provide insight into the associations between these factors and immune subpopulation characteristics. The LONMF factor 13 is predominantly driven by CD56, with a low contribution from CD16, and its high weight coincide with low CD16 and high CD56 characteristics [34], fully reflecting the biology of CD56 bright NK cells (Fig. 6a). Among the 50 factors, factor 13 is the only one with a high Pearson correlation with CD56 and a low correlation with CD16, while CD56 ranks first among the top genes of factor 13 (see Supplementary Figs. S6, S7, S8 of Supplementary Materials 1). Conventional dendritic cells (cDC) are characterized by high expression of FceRI [36]. LONMF factor 44 is driven by FceRI, and its high weight is highly consistent with high FceRI expression, and it is the only factor with a high Pearson’s correlation with FceRI (Fig. 6a). Plasmacytoid dendritic cells (pDC) are phenotypically characterized as CD304-positive and CD14-negative [37]. LONMF factor 35, which is primarily driven by CD304 and contributes less to CD14, is highly consistent in its high weighting with high CD304, demonstrating a strong correlation with pDC biological characteristics (Fig. 6a). Naive CD4 T cells are characterized by high CD4 expression and higher levels of CD45RA than CD45RO [38]. LONMF factor 34 is driven by CD4, and its high weight reflect the expression characteristics of high CD4, high CD45RA, and low CD45RO simultaneously, which is highly consistent with the biological characteristics of naive CD4 T cells (Fig. 6a). Analysis of the Liu dataset. a. Comparison of protein weights and ADT profiles of the four LONMF factors. b. The top genes, proteins, gene sets, and transcription factors (TFs) associated with the 4 factors are displayed in the panel Figure 6b shows a UMAP plot of LONMF, which visualizes the similarity between factor weight distributions and their associated immune subpopulation protein marker activities. Supplementary Fig. 6 shows UMAP visualization results for all factors and marker proteins. LONMF identifies key genes, proteins, and transcription factors in several immune cell subpopulations. For example, in CD56dim NK cells (factor 12), genes such as NCAM1 (CD56), KLRF1, and GZMB, which are associated with NK cell function, and EOMES, TBX21, and MAFF, which regulate differentiation and effector functions, were detected [39]. For native B cells (factor 43), LONMF recognizes genes such as FCER2 (CD23), classical B cell markers such as CD19 and CD21, as well as EBF3 and NF-B (REL/RELA), which are closely related to development and function [40]. For memory B cells (factor 44), top genes include IGHA1 and IGKC, top proteins include CD19 and HLA-DR, and transcription factors associated with their differentiation and homing include RELA, TCF4, and MAX::MYC [41]. Finally, in effector memory CD8 T cells (Factor 39), LONMF extracts CRTAM associated with cytotoxicity, the memory T cell marker CD45RO, and the key transcription factor EOMES [42]. These results demonstrate that LONMF is able to accurately capture the biological characteristics of different immune subpopulations. Conclusion and discussion Currently, several technologies have enabled multi-omics analysis of the same set of cells, and with the rapid development of single-cell multi-omics sequencing technologies, the volume of multi-omics data from a single cell continues to increase, further emphasizing the need and urgency to develop data integration methods. Here, we propose LONMF, a non-negative matrix factorization algorithm based on graph Laplacian and optimal transfer, to integrate different sets of cell data. We thoroughly evaluate the performance of LONMF in cell embedding and clustering tasks by selecting the PBMC 10 × dataset, the CITE-seq dataset, and the scCAT-seq dataset of cancer cell lines for testing, and performing a comprehensive comparative analysis using metrics such as ARI, silhouette coefficient, and purity coefficient. The results show that LONMF outperforms the other four methods (Mowgli, Cobolt, MOFA +, and NMF) in terms of overall performance. This further demonstrates that LONMF is able to effectively preserve the high-dimensional structure of multi-omics datasets while providing significant advantages in the integrated analysis of multi-omics data. In addition, LONMF demonstrates excellent potential in controlled experiments on cancer cell line scCAT-seq datasets, even under challenging conditions. Second, LONMF shows good interpretability. In the PBMC 10 × dataset (RNA + ATAC), in-depth analysis of Factor 8 and Factor 10 revealed that these two factors are highly correlated with immune cells CD56 bright NK cells and memory B cells, respectively. Further Gene Ontology (GO) Biological Process (BP) enrichment analysis of these two factors shows that they are highly correlated with leukocyte-mediated immunity and immune response-regulated cell surface receptor signaling. The results show that they are closely related to key biological processes such as “leukocyte-mediated immunity” and “immune response-regulated cell surface receptor signaling”. Additionally, we evaluate LONMF’s performance on TEA-seq data, demonstrating its capability to establish one-to-one correlations between factors and cell types. The method successfully captures the unique biological characteristics of NK cells, CD8 T cells, CD4 T cells, and others. Furthermore, LONMF generates biologically meaningful results when applied to paired single-cell RNA sequencing, single-cell ATAC sequencing, and surface protein analysis data. Combined with the clinical significance of CD56, CD16, FceRI, and CD304, among others, the unique characterization of immune cell subpopulations in the immune response is revealed. In particular, the relationship between CD56 and NK cell function, the role of CD16 in antibody-dependent cell-mediated cytotoxic responses, and the contribution of FceRI and CD304 in allergic reactions and antiviral immunity suggest that the combined analysis of these markers contributes to a more comprehensive understanding of the complexity of the immune system. With this approach, key immune cell subpopulations and biomarkers can be accurately identified, providing strong support for the design of personalized protocols for clinical immunotherapy. Meanwhile, the robustness of the hyperparameter is investigated, and the experimental results show that the clustering performance of LONMF remains stable across different instances, a finding that is validated by varying the parameter values. The computational efficiency of LONMF is shown in Supplementary Table S3 of Supplementary Materials 1. Due to its relatively high computational complexity, its operational efficiency is slightly lower compared to other methods. However, LONMF demonstrates excellence in clustering performance and biological interpretability, thereby retaining significant potential value for practical applications and subsequent analyses. However, all benchmarks focusing on paired multi-omics integration face an important limitation, namely the lack of high-quality cell biology annotations. Although fluorescence-activated cell sorting (FACS) is an effective stand-alone annotation method in some cases, such annotations are still scarce in the available paired multi-omics data. In addition, exploring potential regulatory relationships between different histological features can provide new perspectives and directions for model optimization and in-depth analysis. LONMF can also be extended to integrate larger datasets, potentially enhancing analytical accuracy and broadening its application potential across multiple research fields. Supplementary Information Below is the link to the electronic supplementary material. Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Author contributions Conceptualization, M.N. and J.G.; methodology, M.N.; software, M.N.; validation, Y.F., and Q.R.; formal analysis, J.G.; investigation, M.N., and L.W.; resources, M.N.; data curation, M.N., and G.Q.; writing—original draft preparation, M.N.; writing—review and editing, J.G., M.N., Q.R., Y.F., and X.C.; visualization, M.N; supervision, M.N. and J.G.; project administration, J.G.; funding acquisition, J.G. All authors have read and agreed to the published version of the manuscript. Funding This work has been supported by the National Natural Science Foundation of China (Grant No. 12271216, 11831015, and 92370131). Data availability Liu dataset: We obtain a scCAT-seq (RNA + ATAC) dataset from Liu et al. [43], which includes 206 cells derived from three cancer cell lines (HCT116, HeLa-S3, and K562). The dataset can be obtained from the supplementary data in the supplementary information of the original publication and accessed through the following URL: https://www.nature.com/articles/s41467-018-08205-7#Sec22. PBMC10X(RNA + ADT) dataset: We obtain a 10X Genomics Multiome (RNA + ADT) dataset containing 6,478 PBMCs. The dataset is accessible at https://support.10xgenomics.com/single-cell-gene-expression/datasets/3.0.0/pbmc_10k_protein_v3. PBMC10X(RNA + ATAC) dataset: We obtain a 10X Genomics Multiome (RNA + ATAC) dataset containing 9,320 PBMCs. The dataset is accessible at https://www.10xgenomics.com/resources/datasets/pbmc-from-a-healthy-donor-granulocytes-removed-through-cell-sorting-10-k-1-standard-2-0-0. TEA dataset: We access a TEA-seq (RNA + ATAC + ADT) dataset from Swanson et al. [44], which contains 7,084 PBMCs. This dataset can be found under GEO login number GSE158013 at: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE158013. The source code of LONMF is available at https://github.com/NMD-CompBio/LONMF-main.git. Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare no Conflict of interest. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Supplementary Materials Supplementary Material 1 Supplementary Material 2 Supplementary Material 3 Data Availability Statement Liu dataset: We obtain a scCAT-seq (RNA + ATAC) dataset from Liu et al. [43], which includes 206 cells derived from three cancer cell lines (HCT116, HeLa-S3, and K562). The dataset can be obtained from the supplementary data in the supplementary information of the original publication and accessed through the following URL: https://www.nature.com/articles/s41467-018-08205-7#Sec22. PBMC10X(RNA + ADT) dataset: We obtain a 10X Genomics Multiome (RNA + ADT) dataset containing 6,478 PBMCs. The dataset is accessible at https://support.10xgenomics.com/single-cell-gene-expression/datasets/3.0.0/pbmc_10k_protein_v3. PBMC10X(RNA + ATAC) dataset: We obtain a 10X Genomics Multiome (RNA + ATAC) dataset containing 9,320 PBMCs. The dataset is accessible at https://www.10xgenomics.com/resources/datasets/pbmc-from-a-healthy-donor-granulocytes-removed-through-cell-sorting-10-k-1-standard-2-0-0. TEA dataset: We access a TEA-seq (RNA + ATAC + ADT) dataset from Swanson et al. [44], which contains 7,084 PBMCs. This dataset can be found under GEO login number GSE158013 at: https://www.ncbi.nlm.nih.gov/geo/query/acc.cgi?acc=GSE158013. The source code of LONMF is available at https://github.com/NMD-CompBio/LONMF-main.git.