LYNX

Modern spatial biology increasingly profiles the same tissue across several molecular layers, measuring transcripts, proteins, metabolites, and stained histology on adjacent sections. Each layer offers a complementary view of cellular and structural organization, but the modalities differ in spatial resolution, carry technology-specific artifacts, and rarely share a common feature space. Most integration methods sidestep these differences by weighting modalities equally, which blurs the very gradients — zonation, differentiation axes, tumor boundaries — that make spatial data biologically informative.

LYNX addresses this problem by learning a shared latent representation of tissue architecture from paired spatial modalities. Rather than treating a tissue as a discrete set of cell types, it models continuous spatial dynamics, providing a unified coordinate system in which cell-cell interactions, phenotypes, and molecular programs can be traced along smooth gradients. This lets researchers ask how signaling, cell states, and physiological function change across space in both healthy and diseased tissue.

The model was developed by the Azizi Lab in the Department of Biomedical Engineering and the Irving Institute for Cancer Dynamics at Columbia University, in collaboration with the Stockwell Lab. It extends the lab's earlier spatial tools — the reference-free deconvolution model Starfysh and the cell-interaction model AMICI — toward a general framework for multimodal spatial integration.

#Key Features

• Shared latent space across modalities: LYNX embeds transcriptomic, proteomic, metabolomic, and histologic measurements from adjacent sections into a single latent representation, so gradients are expressed in one common coordinate system instead of per-modality frames.
• Continuous gradient modeling: The model resolves microenvironmental gradients as cell-state transitions, signaling pathway shifts, and physiological programs, rather than forcing tissue into discrete clusters.
• Localized interaction inference: It quantifies how cell-cell interactions change along spatial gradients, linking molecular composition to communication patterns at specific tissue locations.
• Robustness to resolution and artifacts: By modeling modality-specific technical differences explicitly, LYNX recovers signal that equal-weighting integration tends to degrade, such as noisy or sparse proteomic channels.
• Cross-tissue generality: The same framework applies across distinct tissues and technologies without per-tissue architectural changes, demonstrated on liver, thymus, and breast tumor samples.

#Technical Details

LYNX is a deep generative model built on a heterogeneous attention-based variational graph autoencoder (HeteroAttnVGAE). Cells or spatial spots are represented as nodes in a spatial graph, and heterogeneous attention lets the encoder weigh contributions from different modalities and neighborhood relationships when inferring the shared latent embedding. The variational formulation yields a probabilistic latent space over which gradients, trajectories, and interaction changes can be computed. The released implementation is written in Python and depends on PyTorch Geometric. Across its reference applications, LYNX integrates spatial transcriptomics with metabolomics in liver, spatial RNA with protein in thymus, and spatial transcriptomics with H&E histology in breast tumor, each pairing joint-measured or adjacent-section modalities.

#Applications

LYNX is aimed at researchers studying how tissue function is organized in space. In liver, it recovers metabolically coupled porto-central remodeling of cell interactions; in thymus, it restores degraded proteomic signal along the cortico-medullary axis; and in breast tumor, it reconstructs branching trajectories toward ductal carcinoma in situ and invasive niches, distinguishing stromal activation states and immune-tumor crosstalk. These capabilities support work on tissue zonation, developmental axes, and the spatial structure of the tumor microenvironment, where mapping continuous gradients matters more than assigning discrete labels.

#Impact

LYNX contributes a unified approach to a fast-growing but fragmented problem: making complementary spatial modalities speak the same language. By framing tissue as a continuous latent space and tying molecular gradients to cell communication, it offers spatial biologists a way to study dynamics that per-modality or equal-weighting pipelines obscure. The work is a preprint awaiting peer review, and code is released under a Creative Commons Attribution-NonCommercial-NoDerivatives license, restricting use to non-commercial settings. As a recent entry in the Azizi Lab's line of spatial models, it broadens their toolkit from deconvolution and interaction inference toward integrated multimodal gradient analysis.