Single-Cell Multi-Omic Integration

The foundational mathematical objective of single-cell multi-omic integration is learning a shared lower-dimensional manifold where cells profiled with different assays occupy proximal positions if they share biological identity.

• Variational autoencoders learn probabilistic embeddings that capture both shared and modality-specific variation
• Canonical correlation analysis (CCA) identifies linear combinations of features that maximize cross-modal correlation
• Anchor-based methods identify mutual nearest neighbors across datasets to guide nonlinear alignment

Example: Seurat v4 uses diagonalized CCA followed by L2-normalization to project scRNA-seq and scATAC-seq into a common space where cell types cluster together regardless of assay origin.

Raw biological data from different assays must be converted into numerical representations suitable for neural network consumption before fusion can occur.

• Gene expression is tokenized as normalized transcript counts per gene
• Chromatin accessibility is binarized or count-based over peak regions
• Protein abundance from CITE-seq uses centered log-ratio transformed counts
• DNA methylation is represented as beta values or methylation fractions per CpG site

Each modality requires its own encoder subnetwork that accounts for the statistical properties of that data type—zero-inflation in scRNA-seq, sparsity in scATAC-seq, and compositional constraints in protein data.

Transformer-based architectures use cross-attention layers to allow one modality to selectively query information from another, enabling context-dependent integration.

• A gene expression token can attend to chromatin accessibility peaks near its transcription start site
• Attention weights become interpretable, revealing which regulatory elements drive expression of specific genes
• Multi-head attention captures different types of cross-modal relationships simultaneously

Example: scGPT employs a unified transformer where modality-specific tokens interact through self-attention, learning that an open promoter in ATAC-seq data should influence the expression value of its corresponding gene in RNA-seq data.

A critical capability of multi-omic integration models is computationally predicting absent data layers from observed ones, addressing the reality that most single-cell experiments measure only one or two modalities.

• TotalVI jointly models RNA and protein data, enabling protein prediction from transcriptome alone
• MultiVI extends this to three modalities: RNA, ATAC, and protein
• Generative models learn the joint distribution P(RNA, ATAC, Protein) and can sample from conditional distributions P(Protein | RNA)

This enables in silico multi-omics where existing single-modality datasets gain predicted additional layers, dramatically expanding the utility of legacy data.

Technical variation from different laboratories, protocols, and sequencing platforms creates systematic biases that can overwhelm biological signal during integration.

• Adversarial training penalizes the model when batch identity can be decoded from the latent space
• Harmony iteratively corrects embeddings by clustering and applying batch-specific linear adjustments
• scVI models batch effects as learnable parameters in a hierarchical Bayesian framework

Effective correction ensures that cells cluster by biological identity rather than by which institution generated the data, enabling true multi-study meta-analysis.

A shared, lower-dimensional mathematical representation where embeddings from distinct biological modalities (e.g., scRNA-seq and scATAC-seq) are aligned. This space enables cross-modal comparison by positioning semantically similar biological states—such as a specific cell type—in proximal positions, regardless of which assay generated the data. The quality of single-cell multi-omic integration is fundamentally measured by the coherence and mixing of modalities within this latent representation.

The computational process of mapping feature vectors from different biological assays into a common coordinate system. Techniques range from Canonical Correlation Analysis (CCA) to contrastive learning objectives. The goal is to ensure that a T-cell identified via RNA expression occupies the same latent neighborhood as the same T-cell identified via chromatin accessibility, effectively translating between the languages of transcriptomics and epigenomics.

A fusion technique using attention mechanisms to dynamically weigh the importance of different omics layers for a specific prediction task. For instance, when predicting drug response, the model might learn to prioritize proteomics data over methylation data. This is implemented via cross-attention layers where one modality (e.g., gene expression) queries contextual information from another (e.g., chromatin state).

A regularization technique where entire data modalities (e.g., DNA methylation) are randomly zeroed out during training. This forces the model to learn robust representations that do not over-rely on any single assay. Critically, it prepares the model for real-world clinical deployment where certain omics layers are frequently missing due to cost or tissue limitations, enabling missing modality imputation.

A statistical framework that decomposes the variance of multiple omics datasets into a low-rank matrix of latent factors. MOFA reveals the principal sources of biological and technical variation, distinguishing between factors that drive coordinated changes across all modalities and those that capture modality-specific noise. It is a foundational tool for unsupervised multi-omic exploration.