Single-Cell Multi-Omics Integration: Methods and Applications

Single-cell multi-omics integration combines data from multiple molecular layers measured within individual cells, such as RNA expression, chromatin accessibility, protein abundance, and DNA methylation, to resolve cellular states that cannot be identified from any single modality alone. For researchers working with scRNA-seq and scATAC-seq data, the central challenge is selecting an integration strategy that matches the data structure, the biological question, and the computational resources available. This article provides a decision framework for choosing among integration tools based on whether your data are paired or unpaired, whether you need to predict missing modalities, and whether your goal is cell typing, trajectory inference, or gene regulatory network reconstruction.

The Rationale for Measuring Multiple Molecular Layers in Single Cells

Joint analysis of the genome, epigenome, transcriptome, proteome, and metabolome from single cells is transforming understanding of cell biology in health and disease, according to a review in Nature Reviews Genetics [5]. In less than a decade, technological revolutions have enabled new insights into the interplay between intracellular and intercellular molecular mechanisms that govern development, physiology, and pathogenesis [5]. The core rationale is that each molecular layer provides a partial view of cellular state. Transcript abundance indicates what genes are being expressed, chromatin accessibility reveals which regulatory regions are open for transcription factor binding, and surface protein levels reflect post-transcriptional regulation that RNA measurements cannot capture.

Single-cell multi-omics technologies characterize cell states by simultaneously integrating various single-modality omics methods that profile the transcriptome, genome, epigenome, epitranscriptome, proteome, metabolome, and other emerging omics, as described in Nature Reviews Molecular Cell Biology [6]. These methods have been adapted and improved over the past decade using a framework characterized by optimization of throughput and resolution, modality integration, uniqueness, and accuracy [6]. The impact of these technologies spans cell lineage tracing, tissue-specific and cell-specific atlas production, tumor immunology, cancer genetics, and mapping of cellular spatial information in fundamental and translational research [6].

For a typical research project, the decision to generate multi-omics data should be driven by a specific biological question that cannot be answered with a single assay. For example, linking chromatin accessibility changes to gene expression changes in the same cell can reveal whether an open chromatin region is functionally associated with transcriptional activation. Measuring surface proteins alongside transcripts can distinguish cell populations that are transcriptionally similar but phenotypically distinct. The cost and complexity of multi-omics experiments are justified when the integration of layers provides mechanistic insight that separate measurements cannot.

Core Principles of Multi-Omics Data Integration

Data Structures and Modality Relationships

The problem of integrating different omics data with very different dimensionality and statistical properties remains challenging, as noted in Genomics, Proteomics & Bioinformatics [8]. A growing body of computational tools has been developed for this task, leveraging ideas ranging from machine translation to the theory of networks [8]. Understanding the data structure is the first step in selecting an appropriate method.

Paired data, also called matched data, means that multiple modalities are measured from the same physical cell. Examples include CITE-seq, where RNA and surface proteins are measured simultaneously, and 10x Multiome, where RNA and chromatin accessibility are measured from the same nucleus. Unpaired data means that different modalities are measured from different cells, often from the same tissue sample but not the same individual cell. Integration of unpaired data requires aligning cells across modalities based on shared biological state instead of direct correspondence.

Horizontal integration refers to combining datasets that share the same modalities across multiple batches or conditions. Vertical integration refers to combining different modalities measured from the same cells. Mosaic integration involves datasets where some cells have one modality and other cells have a different modality, with partial overlap. The distinction between these integration scenarios determines which algorithms are appropriate, as demonstrated in a benchmark study published in Nature Methods [11].

The Integration Task Determines the Method

Integration methods can be categorized by the task they perform. Dimension reduction seeks a shared low-dimensional representation of cells across modalities. Batch correction removes technical variation while preserving biological variation. Cell type classification and clustering assign cells to populations based on integrated signals. Imputation predicts missing modality values for cells that lack a particular measurement. Feature selection identifies informative genes, peaks, or proteins. Spatial registration aligns multi-omics data with spatial transcriptomic coordinates [19].

A multitask benchmarking study published in Nature Methods emphasizes that navigating and selecting the most pertinent integration approach poses a significant challenge, contingent upon the tasks relevant to the study goals and the combination of modalities and batches present in the data at hand [19]. The study develops a guideline for choosing the most appropriate method through systematic categorization and comprehensive benchmarking of current methods [19]. This means that researchers should define their downstream analysis goals before selecting an integration tool, instead of applying a default method.

At a Glance: Selecting an Integration Approach

The following table summarizes decision criteria for common single-cell multi-omics integration scenarios. These recommendations are based on published benchmark studies and methodological reviews.

Technologies for Generating Single-Cell Multi-Omics Data

Transcriptome Plus Epigenome

The most widely used paired single-cell multi-omics technologies combine RNA expression with chromatin accessibility. These platforms measure open chromatin regions and gene expression from the same nucleus, enabling direct linkage between regulatory element activity and transcriptional output. The joint analysis of these layers supports gene regulatory network inference, as transcription factor binding sites in open chromatin can be correlated with target gene expression in the same cell [9].

Transcriptome Plus Proteome

CITE-seq and similar technologies measure surface protein abundance alongside transcript counts using oligonucleotide-conjugated antibodies. The statistical problem is that the two views have different dimensions, noise models, and dynamic ranges, yet downstream analysis requires a common representation [16]. Transcript counts are sparse and count-based, while protein abundances are more continuous and have different detection limits. Integration methods must account for these differences to produce meaningful joint representations.

Emerging Modalities

Beyond RNA, chromatin, and protein, single-cell technologies now profile DNA methylation, histone modifications, and spatial information. The field continues to expand with new methods that add modalities or improve throughput and resolution [6]. Spatial multi-omics technologies add a geographic dimension, enabling mapping of molecular layers within tissue architecture [5]. These emerging technologies create new integration challenges because each modality has unique statistical properties and noise characteristics.

Computational Workflow for Multi-Omics Integration

Step 1: Define the Biological Question and Data Structure

Before processing data, document the experimental design and the integration task. Record whether the data are paired or unpaired, which modalities are present, how many batches or samples are included, and what downstream analyses are planned. This information determines the class of integration methods that are appropriate. A study focused on cell type annotation in a well-characterized tissue may require only a simple shared representation, while a study of regulatory mechanisms in a rare cell population may require sophisticated GRN inference.

Step 2: Quality Control and Preprocessing

Each modality requires modality-specific quality control before integration. For scRNA-seq, filter cells based on library size, number of detected genes, and mitochondrial read fraction. For scATAC-seq, filter cells based on total fragments, fraction of fragments in peaks, and transcription start site enrichment. For protein data, assess background signal and batch effects. The quality control thresholds should be recorded and reported for reproducibility.

After quality control, normalize each modality appropriately. RNA data typically use log-normalization or variance-stabilizing transformations. Chromatin accessibility data require term frequency-inverse document frequency normalization or similar approaches that account for sequencing depth differences. Protein data may require arcsinh transformation or centered log-ratio normalization. Applying the wrong normalization to a modality can introduce artifacts that propagate through integration.

Step 3: Select Integration Method Based on Data Structure

Use the decision table above to narrow the candidate methods. For paired data with a clear vertical integration task, methods such as Seurat, MOJITOO, and scAI have demonstrated strong performance [11]. For horizontal integration across batches, totalVI and UINMF are leading choices [11]. For prediction tasks where one modality must be imputed from another, totalVI and scArches for protein prediction and LS_Lab for chromatin accessibility prediction are supported by benchmark evidence [11].

For large-scale datasets, consider whether the method scales to millions of cells. Some methods require substantial memory and compute resources. Foundation models such as scMamba are designed to handle large-scale atlases without prior feature selection [22]. The choice between a classical method and a deep learning method should consider the dataset size, available compute, and the need for interpretability.

Step 4: Run Integration and Assess Quality

After running an integration method, assess the quality of the joint representation using multiple criteria. Check whether known cell types form coherent clusters in the integrated space. Evaluate batch mixing by assessing whether cells from different batches are intermingled within the same biological cell type. Assess modality alignment by checking whether paired cells from the same biological state map to nearby positions in the shared representation.

Benchmark studies provide guidance on expected performance. A Nature Methods benchmark evaluated 14 protein abundance and chromatin accessibility prediction algorithms and 18 integration algorithms using 47 single-cell multi-omics datasets [11]. The study also provides a pipeline to assist researchers in selecting the optimal prediction and integration algorithm [11]. Using such benchmarks to set expectations can help identify integration failures early.

Step 5: Validate Biological Findings

Integration results should be validated using independent evidence. If integration reveals a novel cell population, validate with orthogonal assays such as immunohistochemistry or flow cytometry. If integration identifies a regulatory relationship between chromatin accessibility and gene expression, validate with perturbation experiments. Computational integration generates hypotheses, but experimental validation is required to confirm biological significance.

Options and Tradeoffs in Integration Methods

Classical Statistical Methods

Classical methods include canonical correlation analysis, non-negative matrix factorization, and probabilistic models. These methods are often more interpretable than deep learning approaches because the latent factors can be directly associated with genes, peaks, or proteins. scMFG, for example, uses feature grouping and group integration to maintain interpretability while accurately identifying cell types [21]. The method organizes features with similar characteristics within each omics layer and ensures consistent feature grouping across layers [21].

The tradeoff is that classical methods may not capture complex nonlinear relationships between modalities. They also may require careful parameter tuning and may not scale to very large datasets. However, for many research questions, the interpretability advantage outweighs the performance difference.

Deep Learning Methods

Deep learning methods use neural networks to learn shared representations across modalities. Multimodal deep learning approaches have shown better performance over classical computational methods in integrating multi-omics data, according to a review in Briefings in Bioinformatics [12]. The review categorized deep learning-based integration methods according to data modality, deep learning architecture, fusion strategy, key tasks, and downstream analysis [12].

Deep learning methods can capture nonlinear relationships and scale to large datasets. However, they require substantial computational resources and training data. Interpretability is a persistent challenge, as the learned representations are difficult to associate with specific biological features. Foundation models such as scGPT and scPlantFormer excel in cross-species cell annotation, in silico perturbation modeling, and gene regulatory network inference, according to a review in Journal of Translational Medicine [18].

Graph-Based Methods

Graph-based methods represent cells as nodes in a graph and use graph neural networks to learn representations that preserve local structure. A graph-embedded deep generative clustering model called GeDGC simultaneously learns shared latent representations and cluster factors across multiple omics by leveraging Gaussian mixture models [15]. The model imposes graph embedding constraints on both latent representations and cluster assignments to preserve intrinsic local data structure among cells [15].

The tradeoff is that graph construction requires defining cell-cell similarity, which can be sensitive to the choice of distance metric and neighborhood size. Some methods require additional knowledge to construct cell graphs, which can limit practical utility [15].

Adaptive and Iterative Methods

Recent methods incorporate adaptive strategies to handle heterogeneous data quality across cells. AGSI, an adaptive framework for robust multi-omics integration, uses co-regulated gene modules and iterative reliability assessment [14]. The method identifies co-regulated gene modules using Latent Dirichlet Allocation and evaluates cross-modal correspondence at the module level [14]. Adaptive thresholding dynamically adjusts selection criteria throughout the iterative refinement process [14].

The tradeoff is that iterative methods are computationally intensive and may require careful monitoring to ensure convergence. However, they can achieve superior performance on complex datasets. AGSI achieved up to 25.6% F1 improvement over its ablation baseline and 11.9% over the best competing method on complex neural datasets, and maintained over 85% accuracy even under 50% data dropout [14].

Records and Measurements for Reproducible Integration

Documentation Requirements

Reproducible multi-omics integration requires detailed documentation of every processing step. Record the software versions of all tools used, including the integration method, preprocessing packages, and visualization libraries. Record all parameter values, including normalization methods, dimensionality reduction settings, clustering resolutions, and batch correction parameters. Record the computational environment, including operating system, processor type, and memory allocation.

Record quality control metrics before and after filtering. For each modality, document the number of cells or nuclei passing quality filters, the median number of detected features, and the distribution of key quality metrics. These records enable other researchers to assess the robustness of the integration and to compare results across studies.

Benchmarking Against Published Standards

When possible, benchmark the chosen integration method against published results on similar data types. The Nature Methods benchmark provides a pipeline for selecting optimal prediction and integration algorithms [11]. The multitask benchmarking study in Nature Methods provides a systematic categorization and comprehensive benchmarking of current methods [19]. Using these published benchmarks as reference points can help identify whether the integration performance is within expected ranges.

Data Availability and Sharing

Data sharing is governed by institutional and funder policies. The NIH Genomic Data Sharing Policy outlines expectations for sharing genomic data generated with NIH funding [3]. Researchers should consult their institution and funding agency for specific requirements. The FAIR Guiding Principles provide a framework for making data findable, accessible, interoperable, and reusable [4]. Applying FAIR principles to single-cell multi-omics data includes depositing raw data in public repositories, providing metadata, and using standard file formats.

Common Failure Patterns in Multi-Omics Integration

Modality Imbalance

When one modality has much higher dimensionality or information content than another, integration methods may ignore the weaker modality. For example, RNA data with 20,000 genes can dominate protein data with 200 proteins, leading to integrated representations that primarily reflect transcriptional variation. This failure pattern is detectable by assessing whether the integrated representation preserves modality-specific variation. Methods that use separate branches with modality-specific parameters, such as scCLIP, are designed to retain modality-specific signal through reconstruction [16].

Batch Effects Confounded with Biology

When batch effects correlate with biological conditions, integration methods may remove biological variation along with technical variation. This is particularly problematic in multi-omics studies where different modalities are generated in different batches. Careful experimental design that randomizes samples across batches can reduce this risk. If confounding is unavoidable, document the limitation and consider methods that preserve known biological variation.

Overcorrection

Aggressive batch correction can merge distinct cell populations that are biologically different. This failure pattern is detectable by checking whether known marker genes distinguish populations in the integrated space. If marker gene expression is no longer discriminative after integration, the correction may be too aggressive. Benchmark studies can help calibrate the level of correction appropriate for the data structure.

Feature Selection Bias

Many integration methods rely on selecting highly variable genes or peaks during preprocessing, which may inadvertently discard crucial biological information [22]. This is a particular concern for rare cell populations or subtle biological states that may not exhibit high variability in individual features. Methods that avoid prior feature selection, such as scMamba, address this limitation by preserving genomic positional information [22].

Interpretability Loss

Deep learning methods can achieve high integration performance but produce representations that are difficult to associate with specific biological mechanisms. This is a limitation when the research goal is mechanistic understanding instead of prediction. Methods that maintain interpretability, such as scMFG with its feature grouping approach, can associate outputs with specific cell types or states [21].

Limitations and Interpretation Boundaries

Technical Limitations

Single-cell multi-omics technologies have inherent limitations in sensitivity, throughput, and accuracy [6]. Dropout events, where a feature is not detected despite being present, are common in single-cell data. Integration methods can partially address dropout through imputation, but imputed values should be interpreted with caution. The benchmark study showing that AGSI maintains over 85% accuracy under 50% data dropout indicates that some methods are robust to missing data, but this robustness is not universal [14].

Biological Interpretation Limits

Integration reveals correlations between molecular layers, but correlation does not establish causation. An open chromatin region correlated with gene expression may be a consequence instead of a cause of transcriptional activity. Gene regulatory network inference methods can generate hypotheses about regulatory relationships, but these require experimental validation [9]. The review in Nature Reviews Genetics highlights challenges in GRN inference, particularly with respect to benchmarking [9].

Generalization Limits

Integration methods trained on one tissue or condition may not generalize to others. Cell type definitions vary across tissues, and regulatory relationships are context-specific. Researchers should validate integration results in the specific biological context of their study instead of assuming universal applicability.

Safety and Regulatory Context

Data Privacy and Protection

Single-cell multi-omics data can contain sensitive information about individuals, particularly when combined with clinical data. Researchers must comply with applicable privacy regulations and institutional review board requirements. The NIH Genomic Data Sharing Policy provides guidance on data sharing expectations and protections [3]. Researchers should consult their institution for specific requirements before depositing or sharing data.

Responsible Use of Computational Predictions

Predictions from integration methods, such as imputed protein abundance or predicted chromatin accessibility, should be clearly labeled as computational predictions instead of direct measurements. These predictions are useful for hypothesis generation but should not be used for clinical decision-making without validation. The review in Seminars in Cancer Biology discusses the role of artificial intelligence-based multi-omics analysis in cancer precision medicine, including early screening, diagnosis, response assessment, and prognosis prediction [10]. These applications require rigorous validation before clinical translation.

Professional Escalation Criteria

Researchers should escalate to a bioinformatics core facility, collaborator, or statistical consultant when integration results are inconsistent across methods, when quality control metrics indicate poor data quality, when computational resources are insufficient for the chosen method, or when the biological interpretation of integration results is unclear. Escalation is also appropriate when the research involves human data with privacy concerns or when findings may have clinical implications.

Applications Across Research Domains

Disease Mechanism Discovery

Multi-omics integration has been applied to identify molecular signatures in complex diseases. A study in Frontiers in Immunology integrated bulk RNA-seq, single-cell profiles, spatial transcriptomics, and peripheral blood multi-omics data to uncover key molecular signatures in idiopathic pulmonary fibrosis [7]. The study identified CXCL13, IL33, TLR4, and IGF1 as core genes consistently linked to immune infiltration and fibrotic remodeling [7]. Integration localized their dysregulation to fibroblasts and immune compartments, highlighting fibroblast-immune crosstalk in fibrotic foci [7].

Similarly, a study in Frontiers in Immunology integrated single-cell and bulk transcriptomics to dissect ferroptosis-driven immune remodeling in knee osteoarthritis [20]. The study identified ferroptosis-active homeostasis chondrocytes and revealed ligand-receptor pairs involved in inflammatory cascades [20]. These examples illustrate how integration can reveal cell-type-specific mechanisms that are invisible in bulk data.

Cancer Research

Artificial intelligence-based multi-omics analysis has fueled the development of cancer precision medicine [10]. The coincident development of multi-omics technologies and artificial intelligence algorithms has enabled early cancer screening, diagnosis, response assessment, and prognosis prediction [10]. Single-cell multi-omics technologies have had significant impact in tumor immunology and cancer genetics [6]. The ability to profile the tumor microenvironment at single-cell resolution across multiple molecular layers enables identification of immune cell states, tumor heterogeneity, and potential therapeutic targets.

Gene Regulatory Network Inference

The advent of single-cell multi-omics technologies has led to the development of novel computational methods that leverage genomic, transcriptomic, and chromatin accessibility information to infer gene regulatory networks at unprecedented resolution [9]. These methods can identify transcription factor-gene interactions that are active in specific cell types or states. The review in Nature Reviews Genetics classifies methods that use single-cell multimodal data and highlights challenges in benchmarking [9].

Cell Atlas Construction

Single-cell multi-omics technologies have been instrumental in tissue-specific and cell-specific atlas production [6]. Integration of multiple modalities enables more precise cell type definitions than any single modality. Large-scale atlas projects require integration methods that can handle millions of cells across many batches and donors. Foundation models designed for scalability, such as scMamba, are positioned to handle large-scale atlases [22].

Practical Implementation Steps

Step 1: Inventory Available Data and Resources

Before selecting an integration method, document the available data files, their formats, and the computational resources available. Check whether the data are paired or unpaired, which modalities are present, and whether any modalities are missing for some cells. Assess the total number of cells and the expected computational requirements.

Step 2: Consult Published Benchmarks

Review published benchmark studies to identify methods that perform well for the specific data structure and research goal. The Nature Methods benchmark provides a pipeline for selecting optimal prediction and integration algorithms [11]. The multitask benchmarking study provides guidance on method selection based on tasks and data combinations [19]. These benchmarks can narrow the candidate methods to those with demonstrated performance on similar data.

Step 3: Run Multiple Methods and Compare

Do not rely on a single integration method. Run at least two methods with different underlying assumptions and compare the results. If the methods produce consistent cell type assignments and biological conclusions, confidence in the results increases. If the methods disagree, investigate the source of disagreement. Disagreement may indicate that the data are noisy, that the methods make different assumptions about the data structure, or that the biological signal is weak.

Step 4: Document and Report

Report all integration parameters, quality control metrics, and software versions. Describe the rationale for method selection. Report limitations, including any known batch effects, dropout issues, or interpretability constraints. This documentation enables other researchers to reproduce the analysis and to assess the robustness of the findings.

Frequently Asked Questions

What is the difference between paired and unpaired single-cell multi-omics data?

Paired data means that multiple modalities are measured from the same physical cell, such as RNA and chromatin accessibility from the same nucleus in 10x Multiome experiments. Unpaired data means that different modalities are measured from different cells, often from the same tissue sample. Paired data enable direct linkage between modalities within individual cells, while unpaired data require computational alignment based on shared biological state. The distinction determines which integration methods are appropriate, with vertical integration methods designed for paired data and horizontal or mosaic integration methods designed for unpaired data [11].

How do I choose between classical statistical methods and deep learning methods for integration?

The choice depends on dataset size, computational resources, and the need for interpretability. Classical methods such as matrix factorization and probabilistic models are often more interpretable and require fewer computational resources. Deep learning methods can capture nonlinear relationships and scale to large datasets but are less interpretable and require substantial compute [12]. If the research goal is mechanistic understanding, classical methods may be preferable. If the goal is prediction or handling very large atlases, deep learning methods may be necessary.

What is the best method for predicting protein abundance from single-cell RNA data?

A benchmark study in Nature Methods found that totalVI and scArches outperformed other algorithms for predicting protein abundance from single-cell transcriptomic information [11]. The same study found that LS_Lab was the top-performing algorithm for the prediction of chromatin accessibility in most cases [11]. The choice between these methods should consider the specific data structure and the downstream analysis goals.

How do I know if my integration results are reliable?

Assess integration quality using multiple criteria. Check whether known cell types form coherent clusters in the integrated space. Evaluate batch mixing by assessing whether cells from different batches are intermingled within the same biological cell type. Assess modality alignment by checking whether paired cells from the same biological state map to nearby positions. Run at least two integration methods and compare results. If methods disagree, investigate the source of disagreement. Validate biological findings with orthogonal assays where possible.

What are the common causes of integration failure?

Common failure patterns include modality imbalance, where one modality dominates the integrated representation, batch effects confounded with biology, overcorrection that merges distinct cell populations, feature selection bias that discards important information, and interpretability loss in deep learning methods. Each failure pattern has specific diagnostic indicators and potential remedies. For example, modality imbalance can be detected by assessing whether the integrated representation preserves modality-specific variation, and feature selection bias can be addressed by using methods that avoid prior feature selection [22].

Can integration methods impute missing modalities?

Yes, some integration methods are designed to predict missing modalities. For example, totalVI and scArches can predict protein abundance from RNA data, and LS_Lab can predict chromatin accessibility [11]. However, imputed values should be interpreted with caution. They are computational predictions, not direct measurements, and should be validated experimentally before being used for biological conclusions.

How should I handle batch effects in single-cell multi-omics integration?

Batch effects should be addressed through a combination of experimental design and computational correction. Randomize samples across batches during experimental design to avoid confounding batch with biology. During analysis, use integration methods that include batch correction capabilities. totalVI and UINMF have demonstrated strong performance in horizontal and mosaic integration scenarios that involve multiple batches [11]. Document batch information in the analysis and report any residual batch effects.

What are the data sharing requirements for single-cell multi-omics data?

Data sharing requirements depend on the funding source and institution. The NIH Genomic Data Sharing Policy outlines expectations for sharing genomic data generated with NIH funding [3]. Researchers should consult their institution and funding agency for specific requirements. The FAIR Guiding Principles provide a framework for making data findable, accessible, interoperable, and reusable [4]. Depositing raw data in public repositories and providing detailed metadata supports reproducibility and enables other researchers to build on the work.

Related Bioinformatics Guides

• Single-Cell RNA-Seq Normalization: Batch Effect Correction and Dimension Reduction (PCA, t-SNE, UMAP)
• Multi-Omics Integration Strategies
• Single-Cell ATAC-Seq Bioinformatics
• Single-Cell RNA-Seq Analysis Pipelines for Veterinary Immunology
• Single-cell RNA-seq Trajectory Inference and Cell Lineage Tracing

References and Further Reading

• EMBL-EBI Training. European Bioinformatics Institute.
• NCBI Data Resources. National Center for Biotechnology Information.
• Genomic Data Sharing Policy. National Institutes of Health.
• The FAIR Guiding Principles. Scientific Data.
• Methods and applications for single-cell and spatial multi-omics.. Nature reviews. Genetics, 2023.
• The technological landscape and applications of single-cell multi-omics.. Nature reviews. Molecular cell biology, 2023.
• Multi-omics integration and machine learning reveal gut-immune signatures in idiopathic pulmonary fibrosis: insights from bulk RNA-seq, single-cell profiles, spatial transcriptomics, and experimental validation.. Frontiers in immunology, 2026.
• Computational Methods for Single-cell Multi-omics Integration and Alignment.. Genomics, proteomics & bioinformatics, 2022.
• Gene regulatory network inference in the era of single-cell multi-omics.. Nature reviews. Genetics, 2023.
• Artificial intelligence-based multi-omics analysis fuels cancer precision medicine.. Seminars in cancer biology, 2023.
• Benchmarking algorithms for single-cell multi-omics prediction and integration.. Nature methods, 2024.
• Multimodal deep learning approaches for single-cell multi-omics data integration.. Briefings in bioinformatics, 2023.
• Arsenal of single-cell multi-omics methods expanded.. 2021.
• AGSI: Adaptive group-enhanced strategy for iterative integration of single-cell multi-omics.. 2026.
• Graph-Embedded Deep Generative Clustering for Single-Cell Multi-Omics Data Integration.. 2026.
• scCLIP: A contrastive masked-reconstruction framework for paired single-cell multi-omics integration.. 2026.
• From cellular heterogeneity to precision medicine: single-cell multi-omics in CNS disease research
• Transformative advances in single-cell omics: a comprehensive review of foundation models, multimodal integration and computational ecosystems. Journal of Translational Medicine, 2025.
• Multitask benchmarking of single-cell multimodal omics integration methods. bioRxiv, 2025.
• Single-cell transcriptome and multi-omics integration reveal ferroptosis-driven immune microenvironment remodeling in knee osteoarthritis. Frontiers in Immunology, 2025.
• scMFG: a single-cell multi-omics integration method based on feature grouping. BMC Genomics, 2025.
• scMamba: A Scalable Foundation Model for Single-Cell Multi-Omics Integration Beyond Highly Variable Feature Selection. arXiv.org, 2025.
• The progress and prospects of single-cell sequencing, multi-omics integration analysis, and immunotherapy targeting strategies in lung cancer research. Current Proteomics, 2025.
• scCross: a deep generative model for unifying single-cell multi-omics with seamless integration, cross-modal generation, and in silico exploration. Genome Biology, 2024.

This article is educational and does not replace validated analysis plans, institutional policy, clinical interpretation, or specialist review.