Shared Latent Spaces for Heterogeneous Biological Data

79 concepts74 papers1 proposal

Modern biology has moved beyond the era of simple gene cataloging into what Stuart K. Kim describes as the "modular era," where the goal is to understand how genes, proteins, and metabolites function as coordinated units. However, this transition is hindered by a significant synthesis bottleneck: biological data is fundamentally heterogeneous. As Udo Seiffert et al. note, biological systems operate across scales, producing data that ranges from discrete genomic sequences and categorical clinical markers to continuous gene expression levels and high-dimensional spatial-temporal brain imaging.

The core challenge of heterogeneous data integration is reconciling these structural mismatches. A model cannot simply "add" a protein interaction network to a DNA methylation table; it must find a way to translate these disparate biological languages into a shared mathematical representation.

The Structural Mismatch Problem

The difficulty of integration stems from the diverse formats of biological information. Gal Chechik et al. highlight the friction between linear sequences, relational networks, and continuous expression levels. For example, in single-cell analysis, Felix Drost et al. point out that T-cell receptor data consists of discrete amino acid sequences, while transcriptomics data is continuous and high-dimensional.

Traditional methods often analyze these layers in isolation, which Honghui Yang et al. argue prevents a holistic understanding of how genetic predispositions (like SNPs) manifest as functional abnormalities (like those seen in fMRI scans). To overcome this, researchers have developed three primary strategies to bridge the gap: network fusion, matrix factorization, and deep representation learning.

Bridging Modalities via Network Fusion

One intuitive way to integrate data is to treat every biological entity—whether a gene, a drug, or a patient—as a node in a massive, weighted network. By mapping different data types onto these graphs, researchers can identify "bridge" proteins or hidden functional units.

  • Similarity Networks: Su-Ping Deng et al. demonstrate this by integrating SNPs, DNA methylation, and fMRI data into a single "patient-similarity network." This allows the model to predict disease by looking at how patients cluster across multiple biological layers simultaneously.
  • Topology-Based Prioritization: Yixuan Chen et al. developed the DIR (Data Integration Rank) framework, which merges protein interactions and gene expression into a unified map. By analyzing the global topology of this integrated network, they can identify disease drivers that would be invisible in a single-modality analysis.
  • Steiner Forests: Murodzhon Akhmedov et al. use the "Prize-collecting Steiner Forest" approach to map high-throughput data onto known molecular interaction networks, effectively using the network as a scaffold to synthesize DNA, RNA, and protein expression levels.

Shared Latent Spaces and Factorization

A more abstract but powerful approach involves projecting heterogeneous data into a shared latent space—a mathematical "meeting ground" where different modalities can be compared directly.

Marinka Zitnik et al. pioneered the use of simultaneous matrix tri-factorization for this purpose. By decomposing disparate datasets (like interaction graphs and text-based literature) into shared low-rank components, the model can predict gene functions without requiring the "data cleaning tax" of traditional methods. Similarly, Milla Kibble et al. applied Group Factor Analysis (GFA) to multi-omics data from twins, identifying co-variation across genomics and lifestyle factors by finding the latent factors that explain variance across all data layers.

Deep Representation Learning and Foundation Models

The current frontier of integration leverages deep learning to "translate" biological modalities into vector embeddings. The intuition is that if two different data types (e.g., a protein's sequence and its 3D structure) describe the same biological reality, they should occupy similar positions in a high-dimensional vector space.

  • Cross-Modality Encoding: Sheng Wang 0001 et al. developed ProSNet, which embeds protein sequences and interaction networks into a shared low-dimensional space to predict function.
  • Contrastive Alignment: Peng Yue et al. use contrastive learning to align protein language model embeddings with dynamic experimental "omics" data, allowing the model to incorporate real-world pathogen behavior into sequence-based predictions.
  • Metric Learning: To avoid the "black-box" nature of some deep learning, Rohit Singh et al. introduced Schema, a metric learning framework that reweights features to maximize agreement between modalities (like spatial location and gene expression) while strictly limiting data distortion.

As the field moves toward large-scale foundation models, platforms like PyTDC, developed by Alejandro Velez-Arce et al., are standardizing how we synthesize high-dimensional modalities like 3D structures and genomic sequences. The ultimate goal, as suggested by Shi Qiu et al., is to move from a simple census of biological components to a functional narrative that explains how these heterogeneous layers interact to drive health and disease.

Key Foundational Work

The modular era of functional genomics. (2003) — contributed 2 concept elements to this direction.
Kernel-Based Data Fusion and Its Application to Protein Function Prediction in Yeast (2004) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_heterogeneous_data_integration_heterogen_cb6eb1d4,
  title = {heterogeneous data integration (heterogeneous biological data integration)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/heterogeneous-data-integration-heterogeneous-biological-data-integration/rdr_cb6eb1d48e014954b52efb7019cb8769},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_heterogeneous_data_integration_heterogen_cb6eb1d4_v1,
  title = {heterogeneous data integration (heterogeneous biological data integration)},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/heterogeneous-data-integration-heterogeneous-biological-data-integration/rdr_cb6eb1d48e014954b52efb7019cb8769/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake