🧬 Single-Cell Multi-Omics Integration
A Comprehensive Guide to Methods, Frameworks, and Best Practices (2015-2025)
From early paired measurements to modern foundation models: Understanding how to integrate
RNA-seq, ATAC-seq, protein abundance, spatial data, and perturbation responses at single-cell resolution.
Based on systematic analysis of 40+ landmark papers spanning a decade of innovation.
📖 What is Single-Cell Multi-Omics Integration?
Single-cell multi-omics integration combines measurements of different molecular layers (transcriptome, epigenome, proteome, spatial location) from the same or related cells to build comprehensive cellular maps. This integration is essential for understanding cell states, developmental trajectories, disease mechanisms, and therapeutic responses.
Why Multi-Omics Integration Matters
- Comprehensive Cell State Definition: Single modalities provide incomplete views; RNA tells what's transcribed, ATAC reveals accessible chromatin, proteins show functional output
- Regulatory Mechanism Discovery: Linking chromatin accessibility → transcription → protein abundance reveals gene regulatory networks and signaling cascades
- Batch Effect Correction: Harmonizing data across experiments, technologies, and labs enables atlas-scale analyses and meta-studies
- Missing Modality Imputation: Predicting unmeasured features (e.g., protein from RNA) reduces experimental costs while maintaining biological insights
- Spatial Context Integration: Combining molecular profiles with spatial locations reveals tissue architecture and cell-cell interactions
- Perturbation Response Modeling: Understanding how genetic or chemical perturbations affect multiple molecular layers simultaneously
Integration Paradigms (Fu, Shaliu, et al. Nature Methods, 2025), (Liu, Chunlei, et al. Nature Methods, 2025)
According to the benchmark papers, there are six major multi-omics integration paradigms:
🔵 Vertical Integration
🟢 Diagonal Integration
🟡 Mosaic Integration
🔴 Cross Integration
🟣 Spatial Integration
🟠 Perturbation Integration
Each paradigm addresses different data structures and analytical challenges in multi-omics analysis.
📅 Evolution Timeline: From Paired Measurements to Foundation Models
2015-2017: Early Multi-Modal Technologies
Key Innovations:
- G&T-seq (2015): First simultaneous RNA + DNA methylation
- CITE-seq (2017): RNA + surface protein via antibody tags
- mixOmics (2017): Statistical framework for multi-block data
Era Characteristic: Experimental methods development; simple statistical integration
2019-2020: Statistical Methods Era
Key Innovations:
- DIABLO (2019): Multi-omics discriminant analysis
- MOFA+ (2020): Multi-omics factor analysis with covariates
Era Characteristic: Matrix factorization; interpretable latent factors; limited scalability
2021-2022: Deep Learning Breakthrough
Key Innovations:
- totalVI (2021): VAE for RNA + protein integration
- Seurat WNN (2021): Weighted nearest neighbor multi-modal analysis
- Concerto (2022): Contrastive learning for 10M+ cells
Era Characteristic: VAE dominance; scalability improvements; atlas-scale analyses
2023: Optimal Transport & Graph Methods
Key Innovations:
- CellOT (2023): Neural optimal transport for perturbations
- SIMBA (2023): Graph embedding with cells + features co-embedded
Era Characteristic: Theoretical rigor; optimal transport theory
2024: Foundation Model Era Begins
Key Innovations:
- scGPT (2024): 100M parameter transformer on 33M cells
Era Characteristic: Pre-training paradigm; 10M+ cell datasets; transfer learning
2025: Specialized Foundation Models
Key Innovations:
- CellWhisperer (2025): Instruction-tuned multimodal foundation model
- Nicheformer (2025): Spatial multi-omics foundation model
- OmiCLIP (2025): Visual-omics foundation model (H&E + transcriptomics)
- MORPH (2025): Cross-condition perturbation prediction
Era Characteristic: Task-specific foundation models; comprehensive benchmarking; clinical translation focus
🔬 Method Taxonomy: Algorithmic Approaches
By Computational Framework
🧠 Variational Autoencoders (VAE-based)
Principle: Learn probabilistic latent representations with encoder-decoder architecture
Advantages: Uncertainty quantification; generative modeling; missing data imputation
🔄 Contrastive Learning
Principle: Learn representations by pulling similar samples together, pushing dissimilar apart
Advantages: Scalability to millions of cells; no explicit pairing needed; robust embeddings
Concerto (2022) - 10M+ cells
📊 Graph Neural Networks (GNN)
Principle: Model cells as graph nodes; aggregate information from neighborhoods
Advantages: Captures cell-cell relationships; flexible message passing; spatial awareness
SIMBA (2023) - Cells + features co-embedding
GLUE (2022) - Graph-based integration
SIMVI (2025) - Spatial + intrinsic disentanglement
🚀 Optimal Transport
Principle: Find minimal-cost mapping between cell distributions
Advantages: Theoretical guarantees; preserves distributional structure; interpretable
CellOT (2023) - Perturbation prediction
Labeled GWOT (2025) - Cross-modality alignment
SCOT (2022) - Gromov-Wasserstein
🤖 Foundation Models (Transformers)
Principle: Pre-train large models on massive datasets; fine-tune for specific tasks
Advantages: Transfer learning; few-shot adaptation; generalizable representations
CellWhisperer (2025) - Instruction-tuned
Nicheformer (2025) - Spatial specialist
scGPT (2024) - 33M cells pretrain
OmiCLIP (2025) - Visual-omics CLIP
🔗 Matrix Factorization & Classical
Principle: Decompose data matrices into latent factor representations
Advantages: Interpretable factors; computationally efficient; well-understood theory
MOFA+ (2020) - Multi-omics factors
Seurat CCA/WNN (2021) - Canonical correlation
mixOmics (2017) - Multiblock projection to latent structure (PLS)
DIABLO (2019) - Discriminant analysis
By Scale Capability
Scalability Tiers
- Small Scale (<10K cells): MOFA+, DIABLO, mixOmics - ideal for pilot studies
- Medium Scale (10K-100K cells): Seurat WNN, totalVI, MultiVI - standard analyses
- Large Scale (100K-1M cells): Concerto, SIMBA, scBridge - atlas construction
- Atlas Scale (>1M cells): Foundation models (scGPT, CellWhisperer), SnapATAC2 - population studies
📄 Landmark Papers by Computational Framework (2015-2025)
🧬 Experimental Technologies (Foundation)
CITE-seq: Simultaneous epitope and transcriptome measurement in single cells
Pioneering technology combining RNA-seq with antibody-derived tags (ADT) for protein quantification.
Enabled paired transcriptome-proteome measurements at single-cell resolution.
Key Contributions:
- Antibody-oligonucleotide conjugation method
- Validated on PBMC immune cell populations
- Foundation for multi-modal single-cell biology
scONE-seq: A single-cell multi-omics method enables simultaneous dissection of phenotype and genotype heterogeneity from frozen tumors
scONE-seq is a versatile single-cell multi-omics method that simultaneously profiles whole-genome DNA and full-length RNA from the same cell in a one-pot reaction, enabling multi-omics analysis of frozen biobanked tumor samples and revealing transcriptionally normal-like tumor clones.
Applications:
- Works with frozen tissue samples
- Simultaneous DNA and RNA profiling
- Tumor heterogeneity analysis
🔷 Variational Autoencoders & Probabilistic Models
Probabilistic harmonization and annotation of single-cell transcriptomics data with deep generative models (totalVI)
Variational autoencoder for integrating RNA and protein measurements. Models technical effects
including batch, background noise, and protein zero-inflation.
Capabilities:
- Batch correction across technologies
- Protein imputation from RNA
- Uncertainty quantification
Multi-resolution deconvolution of spatial transcriptomics data reveals continuous patterns of inflammation (MultiVI)
Variational inference framework for mosaic multi-omics integration. Handles incomplete
modality measurements across datasets with joint latent space.
Features:
- Handles RNA+ADT+ATAC combinations
- Missing modality imputation
- Spatial deconvolution capabilities
Cobolt: integrative analysis of multimodal single-cell sequencing data
Cobolt uses Multimodal Variational Autoencoders (MVAE) based on hierarchical generative models to enable coherent integration of multi-modality single-cell data with single-modality datasets, creating a unified representation for downstream analysis.
Innovations:
- Hierarchical latent variable model
- Handles incomplete modality measurements
- Supports SNARE-seq and other multimodal data
🕸️ Graph-Based Methods
Integrated analysis of multimodal single-cell data (Seurat WNN)
Weighted nearest neighbor (WNN) algorithm for integrating paired multi-modal measurements.
Learns modality-specific weights to construct unified cell similarity graph.
Key Innovations:
- Cell-specific modality weighting
- Works with RNA+ADT, RNA+ATAC
- Integrated into widely-used Seurat package
SIMBA: single-cell embedding along with features
Graph-based co-embedding framework placing cells and features (genes, peaks, motifs) in shared space.
Enables clustering-free marker discovery and regulatory network inference.
Innovations:
- Unified cell-feature embedding space
- Multi-omics native support
- Scales to 1.3M cells in 1.5 hours
🔄 Optimal Transport Methods
Learning single-cell perturbation responses using neural optimal transport (CellOT)
Input convex neural networks for optimal transport maps. Predicts single-cell perturbation
responses from unaligned control/treatment populations.
Applications:
- Drug response prediction
- Genetic knockout effects
- Cross-patient generalization
Cross-modality matching and prediction of perturbation responses with labeled Gromov-Wasserstein optimal transport
Label-constrained Gromov-Wasserstein OT for cross-modality alignment. Achieves L-fold
computational speedup while improving alignment quality with perturbation labels.
Innovations:
- Incorporates perturbation labels
- RNA → protein prediction
- Dose-response preservation
🧠 Deep Learning & Neural Networks
Deep learning and alignment of spatially resolved single-cell transcriptomes with Tangram
Tangram is a deep learning framework that aligns single-cell/single-nucleus RNA-seq data to any form of spatial transcriptomics data to generate genome-wide spatial expression maps at single-cell resolution and integrate them with anatomical references.
Key Features:
- Works with any spatial technology (MERFISH, Visium, smFISH)
- Expands gene coverage to genome-wide scale
- Automated histological registration module
scJoint integrates atlas-scale single-cell RNA-seq and ATAC-seq data with transfer learning
scJoint is a scalable transfer learning method using neural networks to integrate atlas-scale scRNA-seq and scATAC-seq data through semisupervised learning, achieving 84% label transfer accuracy while handling over 1 million cells in 2 hours.
Key Achievements:
- Processes 1M+ cells in 2 hours
- 84% label transfer accuracy
- Effective batch correction across platforms
scBridge embraces cell heterogeneity in single-cell RNA-seq and ATAC-seq data integration
scBridge is a heterogeneous transfer learning method that exploits cell heterogeneity to progressively integrate scRNA-seq and scATAC-seq data by first identifying and integrating "reliable" scATAC-seq cells with smaller omics differences, then using them as bridges to integrate the remaining cells.
Novel Approach:
- Progressive integration strategy
- Exploits cell heterogeneity as advantage
- Superior performance on challenging datasets
scMODAL: a general deep learning framework for comprehensive single-cell multi-omics data alignment with feature links
scMODAL is a deep learning framework leveraging neural networks and generative adversarial networks to align single-cell multi-omics datasets with limited known feature correlations, enabling accurate cross-modality integration, feature imputation, and regulatory relationship inference.
Features:
- GAN-based architecture for alignment
- Works with limited feature correlations
- Cross-modality imputation and prediction
A visual-omics foundation model to bridge histopathology with spatial transcriptomics (OmiCLIP)
CLIP-based foundation model trained on 2.2M H&E image-transcriptomics pairs. Enables
gene expression prediction from routine histology images.
Capabilities:
- Image → transcriptomics prediction
- Tissue section alignment
- Cell type annotation from H&E
- Spatial decomposition
🎯 Contrastive Learning
Contrastive learning enables rapid mapping to multimodal single-cell atlas of multimillion scale (Concerto)
Self-distillation framework with asymmetric teacher-student architecture. Demonstrates
linear scalability to 10M+ cells with 100x speedup over baselines.
Achievements:
- 10M cell reference atlas in 1.5 hours
- Query mapping in 8 seconds (10K cells)
- Superior clustering and classification
🤖 Foundation Models & Transformers
scGPT: toward building a foundation model for single-cell multi-omics using generative AI
100M parameter transformer pre-trained on 33M cells. Generative pretraining with cell-as-sentence paradigm.
Supports multi-omics fine-tuning and zero-shot predictions.
Capabilities:
- Cell type annotation
- Batch correction
- Perturbation prediction
- Gene network inference
CellWhisperer: An instruction-tuned foundation model for single-cell multimodal analysis
Instruction-tuned multimodal foundation model supporting natural language queries.
First single-cell model with conversational interface for biological questions.
Features:
- Natural language biological queries
- Multi-task learning (classification, clustering, prediction)
- Zero-shot generalization
🌉 Mosaic & Bridge Integration
Stabilized mosaic single-cell data integration using unshared features (StabMap)
StabMap enables mosaic data integration of single-cell datasets by exploiting non-overlapping features through traversal of a mosaic data topology, allowing multi-hop integration where some datasets share no common features.
Unique Capabilities:
- Multi-hop integration without direct feature overlap
- Leverages non-overlapping features
- Supports supervised and unsupervised modes
Building a cross-species cell atlas with interpretable deep learning (Dictionary Learning)
Dictionary learning approach for cross-species and cross-technology integration. Learns
interpretable gene programs bridging evolutionary distances.
Applications:
- Human-mouse integration
- Cross-platform harmonization
- Conserved program discovery
🔬 Perturbation & Response Prediction
Predicting cell morphological responses to perturbations using generative modeling (IMPA)
IMPA is a generative style-transfer model that predicts cellular morphological responses to unseen chemical and genetic perturbations while accounting for batch effects in high-content imaging screens.
Applications:
- Drug response prediction from imaging
- Handles batch effects in HCS
- Generative modeling for perturbation screens
📊 Benchmark & Review Papers
Multitask benchmarking of single-cell multimodal omics integration methods
Comprehensive evaluation of 40 integration methods across 7 tasks using 64 real + 22 simulated datasets.
Establishes integration taxonomy (vertical/diagonal/mosaic/cross) and provides decision trees for method selection.
Key Findings:
- No universal winner; task-dependent performance
- Deep learning dominates diagonal/cross integration
- Batch correction often trades off with biological preservation
Benchmarking single-cell multi-modal data integrations
Systematic assessment across usability, accuracy, and robustness dimensions. Evaluates 40 algorithms
with 101 benchmark datasets spanning diverse technologies and cell types.
Evaluation Framework:
- Multi-gradient AUC for robustness
- Hardware scalability testing (500GB RAM, 24h limits)
- Cross-modality imputation assessment
How to build the virtual cell with artificial intelligence: Priorities and opportunities (AIVC)
Vision paper outlining the AI Virtual Cell framework. Proposes universal representations across
molecular, cellular, and multicellular scales with virtual instruments for manipulation and decoding.
Framework Components:
- Universal representations (URs) across scales
- Virtual instruments (manipulators & decoders)
- Foundation model architecture for cells
The Human Cell Atlas: from a cell census to a unified foundation model
Roadmap for Human Cell Atlas evolution from cell census to foundation models. Outlines 5 perspectives:
cell census, 3D maps, genotype-phenotype maps, developmental maps, and foundation models.
🔧 Methods Development & Innovation
SnapATAC2: A fast, scalable and versatile tool for analysis of single-cell omics data
Matrix-free spectral embedding for chromatin accessibility data. Achieves linear time/memory
complexity enabling analysis of million-cell datasets on standard hardware.
Performance:
- O(n) complexity vs O(n²) for competitors
- 63.4% cost reduction vs ArchR
- 200K cells in 13.4 minutes
SIMVI disentangles intrinsic and spatial-induced cellular states in spatial omics data
Dual encoder VAE with MLP for intrinsic variation and GAT for spatial variation. First
identifiable framework for separating intrinsic cell state from spatial microenvironment effects.
Innovations:
- Asymmetric regularization for identifiability
- Single-cell spatial effect estimation
- Causal inference integration (DML)
scMODAL: a general deep learning framework for comprehensive single-cell multi-omics data alignment
GAN-based integration using feature links (e.g., gene-protein pairs). First method to effectively
handle weakly-linked modalities through adversarial alignment.
Capabilities:
- 29-34% imputation improvement
- Works with minimal feature links
- Integrates CITE-seq + CyTOF
MORPH predicts the single-cell outcome of genetic perturbations across conditions and data modalities
Discrepancy-based VAE with attention mechanism for perturbation prediction. Handles both
transcriptomic and imaging modalities with prior knowledge integration (DepMap, GenePT).
Features:
- Cross-cell line transfer learning
- Combinatorial perturbation modeling
- Active learning for experiment design
MetaQ: fast, scalable and accurate metacell inference via single-cell quantization
Deep learning with quantization for metacell construction. Achieves O(n) complexity and
100x speedup over SEACell while maintaining superior classification accuracy.
Performance:
- 100K cells in 0.3h vs 26.7h (SEACell)
- 88% balanced accuracy vs 84% (baseline)
- Native multi-omics support
ADTnorm: robust integration of single-cell protein measurement across CITE-seq datasets
Functional data analysis for protein expression normalization. Landmark alignment preserves
negative peaks essential for cell type annotation while harmonizing batch effects.
Applications:
- Cross-institutional integration
- Antibody titration optimization
- Auto-gating (80-100% accuracy)
📚 Comprehensive Method Comparison
By Integration Category & Performance
| Method | Year | Category | Modalities | Scale | Key Strength |
|---|---|---|---|---|---|
| Vertical Integration (Paired Multi-Modal) | |||||
| Seurat WNN | 2021 | Vertical | RNA+ADT, RNA+ATAC | ~100K cells | Cell-specific modality weighting; widely adopted |
| totalVI | 2021 | Vertical/Cross | RNA+ADT | ~50K cells | Probabilistic; batch correction; imputation |
| Multigrate | 2024 | Vertical/Cross | RNA+ADT+ATAC | ~100K cells | Tri-modal support; robust performance |
| Diagonal Integration (Unpaired, Non-Overlapping) | |||||
| scBridge | 2023 | Diagonal | RNA+ATAC | ~50K cells | Superior dimensionality reduction & clustering |
| GLUE | 2022 | Diagonal | RNA+ATAC | ~50K cells | Graph neural network; best batch correction |
| scJoint | 2022 | Diagonal | RNA+ATAC | ~100K cells | Multi-batch integration; transfer learning |
| Mosaic Integration (Overlapping Incomplete) | |||||
| StabMap | 2023 | Mosaic | Any combination | ~50K cells | Flexible; efficient; handles any modality pattern |
| MultiVI | 2023 | Mosaic | RNA+ADT+ATAC | ~100K cells | VAE-based; missing modality imputation |
| Cobolt | 2023 | Mosaic | RNA+ADT+ATAC | ~50K cells | Bayesian framework; uncertainty quantification |
| Spatial Integration | |||||
| SIMVI | 2025 | Spatial | Spatial transcriptomics | ~60K cells | Disentangles intrinsic vs spatial variation |
| OmiCLIP | 2025 | Spatial | H&E + ST | 2.2M pairs | Visual-omics foundation model; H&E → gene expression |
| Tangram | 2021 | Spatial | Spatial mapping | ~50K cells | Maps scRNA-seq to spatial coordinates |
| Perturbation-Aware Integration | |||||
| CellOT | 2023 | Perturbation | RNA-seq (protein/imaging) | ~50K cells | Neural OT; single-cell predictions |
| MORPH | 2025 | Perturbation | RNA + Imaging | ~300K cells | Cross-modality; cross-cell line transfer |
| Labeled GWOT | 2025 | Perturbation | RNA + Protein | ~50K cells | Label-constrained OT; L-fold speedup |
| Classical/Statistical Methods | |||||
| MOFA+ | 2020 | Vertical | Any | ~10K cells | Interpretable factors; handles covariates |
| mixOmics | 2017 | Vertical | Any | ~5K cells | Multiblock projection to latent structure (PLS); statistical rigor |
💡 Practical Implementation Guide
Choosing the Right Method: Decision Framework
Step 1: Identify Your Data Structure
- All cells have all modalities? → Vertical integration
- Different batches, different modalities? → Diagonal
- Mixed modality availability? → Mosaic
- Multiple batches, all modalities? → Cross
- Spatial data? → Spatial methods
- Perturbation data? → Perturbation-aware
Step 2: Consider Your Computational Resources
Limited Resources (no GPU, <32GB RAM):
- Classical methods
Moderate Resources (GPU optional, 32-64GB RAM):
- Most VAE-based and graph methods
High-End Resources (GPU required, 64GB+ RAM):
- Foundation models
Common Pitfalls & Best Practices
- Data Quality First: Ensure proper preprocessing (UMI counts >500, doublet removal, quality control). Use standard pipelines (Scanpy, Seurat). Bad data → bad integration.
- Modality-Specific Normalization: Each modality needs appropriate normalization (RNA: log-normalization, ATAC: TF-IDF, Protein: ADTnorm/DSB). Don't use RNA normalization for ATAC!
- Evaluation Metrics Matter: Use multiple complementary metrics (ARI, NMI, ASW for clustering; kBET, iLISI for batch mixing). Single metrics can be misleading.
- Batch Effect vs Biological Variation: Over-correction removes biology; under-correction leaves technical artifacts. Check cell type separation alongside batch mixing.
- Held-Out Testing: Evaluate on held-out batches or cell types, not just held-out cells from training conditions. Tests true generalization.
- Baseline Comparisons: Always compare against simple baselines (PCA, naive concatenation, per-modality analysis). Complex methods should outperform simple ones.
- Computational Scalability: Test methods on your actual data size. Many methods fail silently or degrade performance on large datasets.
- Biological Validation: Computational integration should reveal biologically meaningful patterns. Validate key findings with marker genes, known cell types, or experimental follow-up.
- Hyperparameter Tuning: Default parameters often suboptimal. Use cross-validation or grid search for critical parameters (latent dimensions, regularization, learning rate).
- Reproducibility: Report random seeds, package versions, preprocessing steps, and hyperparameters. Provide code and (when possible) data.
Software Ecosystem & Tools
- Scanpy: Standard Python toolkit for preprocessing, analysis, visualization. Foundation for many integration methods.
- Seurat: R package with Seurat v3/v5 integration, WNN, UMAP. Industry standard for single-cell RNA-seq.
- scvi-tools: PyTorch-based framework for probabilistic models (scVI, totalVI, MultiVI, scANVI). GPU-accelerated.
- Muon: Python package for multimodal omics analysis. Integrates with Scanpy ecosystem.
- LIGER: R package for integrative non-negative matrix factorization. Good for cross-platform integration.
- Harmony: Fast batch correction algorithm. Works on PCA embeddings. Widely used for large-scale integration.