Shared Latent Spaces for Multi-Omic Data Integration

329 concepts304 papers4 proposals

Cells are dynamic systems governed by the flow of information from DNA to RNA to proteins. Yet, high-throughput sequencing technologies typically capture only a single slice of this reality at a time. A transcriptomic snapshot reveals mRNA abundance but ignores chromatin accessibility, spatial organization, or protein translation kinetics.

To build a holistic model of cellular state, we must integrate these disparate modalities. The fundamental challenge is that each "omic" layer speaks a different mathematical language: RNA data is high-dimensional and zero-inflated, protein data is dense but limited in scope, and epigenetic data is sparse and binary. Multi-omic data integration solves this by finding the shared biological truth hidden beneath these different data structures.

Aligning Modalities via Shared Latent Spaces

Why does multi-omic integration work? It works because different molecular layers are not independent; they are constrained by the underlying biophysical state of the cell. Deep generative models exploit these constraints by mapping heterogeneous data types into a shared, low-dimensional latent space.

In single-modality settings, architectures like the single-cell Embedded Topic Model (scETM), developed by Huiyu Cai et al. in Learning interpretable cellular and gene signature embeddings from single-cell transcriptomic data, demonstrate that linear decoders can map transcriptomic data into interpretable embeddings linked directly to biological pathways. Similarly, the DREAM framework by Jing Jiang et al. in Dimensionality reduction and visualization of single-cell RNA-seq data with an improved deep variational autoencoder uses a variational autoencoder with a zero-inflated layer to filter out technical noise and capture true biological variance.

When extending this intuition to multi-omics, the goal is to force different encoders (e.g., one for RNA, one for DNA) to project their respective data into the same latent coordinates. Once aligned, a distance metric in this shared space reflects true functional differences rather than modality-specific artifacts.

[Modality A: RNA] ---> (Encoder A) ---\
                                       ---> [ Shared Latent Space ]
[Modality B: DNA] ---> (Encoder B) ---/

This alignment does not always require measuring every modality in the exact same physical cell. For instance, Kieran R. Campbell et al. developed clonealign (clonealign: statistical integration of independent single-cell RNA and DNA sequencing data from human cancers), which statistically links independent single-cell RNA and DNA sequencing data. It works by leveraging the "dosage effect"—the biological correlation between gene copy number and expression levels. By anchoring the mathematical alignment on known biological relationships, the model maps independent modalities into a unified representation.

Cross-Modal Translation and Foundation Models

If a shared latent space captures the joint distribution of multiple modalities, we can treat multi-omic integration as a translation problem. Given an easily accessible modality (like standard scRNA-seq), we can predict a more expensive or destructive measurement (like proteomics).

This is the intuition behind scTEL, a framework designed by Yuanyuan Chen et al. in A joint analysis of single cell transcriptomics and proteomics using transformer. By using a Transformer-LSTM architecture, scTEL learns the complex, non-linear mapping from the transcriptome to the proteome. This translation works because the model implicitly captures the cellular translation machinery and degradation kinetics that link mRNA to protein abundance, allowing researchers to gain multi-omic insights from single-modality data.

On a larger scale, biological foundation models are moving toward multi-modal pretraining. Early models like HyenaDNA by Eric Nguyen et al. (HyenaDNA: Long-Range Genomic Sequence Modeling at Single Nucleotide Resolution) focused strictly on processing long-range DNA sequences at single-nucleotide resolution. However, as argued by Christopher Blum et al. in Multi-modal Transfer Learning between Biological Foundation Models, the future lies in flexible architectures that can simultaneously process and relate DNA, RNA, and protein structures to exploit the complementary signals embedded within the central dogma.

Grounding Alignment with Spatial and Prior Knowledge

Unsupervised alignment of high-dimensional spaces is prone to overfitting; mathematical noise can easily be mistaken for biological signal. Modern frameworks solve this by grounding the alignment using two anchors: physical space and prior biological knowledge.

Physical Space as an Anchor

Cells do not exist in isolation; their microenvironment dictates their state. Frameworks like stMSA by Han Shu et al. (Efficient integration of spatial omics data for joint domain detection, matching, and alignment with stMSA) use graph autoencoders and contrastive learning to integrate spatial coordinates with transcriptomic and proteomic data, enabling precise 3D tissue reconstruction.

By treating spatial coordinates as a structural constraint—or even as a "language" as seen in the Spatial2Sentence framework by Chi-Jane Chen et al. (Spatial Coordinates as a Cell Language: A Multi-Sentence Framework for Imaging Mass Cytometry Analysis)—models leverage neighbor interactions to resolve ambiguities in molecular data. If two cells have ambiguous transcriptomic profiles, their spatial context can act as the tie-breaker for alignment.

Prior Knowledge as a Regularizer

Instead of forcing a neural network to learn everything from scratch, researchers inject known biological structures directly into the network topology. In Combining gene ontology with deep neural networks to enhance the clustering of single cell RNA-Seq data, Jiajie Peng proposed GOAE and GONN, which replace standard black-box layers with partially connected layers derived from Gene Ontology (GO) terms.

Because the network's connections mirror actual biological pathways, the model is constrained to learn representations that are inherently interpretable and robust to noise. This architectural constraint ensures that when multi-omic signals are fused, they converge on verified, functional cellular mechanisms rather than statistical artifacts.

History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_multi_omic_data_integration_8948a278,
  title = {multi-omic data integration},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-omic-data-integration/rdr_8948a27835b7404f98e4048f9315f9d5},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_multi_omic_data_integration_8948a278_v1,
  title = {multi-omic data integration},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/multi-omic-data-integration/rdr_8948a27835b7404f98e4048f9315f9d5/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake