Generative diffusion model that reassembles dissociated single cells into tissues, predicting spatial coordinates from gene expression alone.
No providers recorded yet. Browse all providers
Single-cell RNA sequencing measures expression at the price of position: dissociating a tissue destroys the spatial arrangement that determines which cells signal to which. Spatially resolved technologies preserve that arrangement but cover fewer genes, fewer cells, or fewer tissues, and vast archives of existing scRNA-seq data have no spatial counterpart at all. LUNA, from the Brbić lab at EPFL, attacks the problem from the other direction: it learns spatial priors from tissues that have been profiled spatially, then generates coordinates for dissociated cells given nothing but their gene expression.
LUNA is a generative diffusion model over cell positions. Given a set of expression profiles, it denoises a random spatial configuration into one consistent with the tissue organisation it learned during training. Crucially the model is conditioned only on expression — no cell-type labels, no reference image, no anchor spots — so it can be applied to atlases collected long before spatial assays existed.
The model shares its goal with catalog entries such as Nicheformer, which also transfers spatial context onto dissociated cells, but frames reconstruction as generative denoising of cell-cell geometry rather than as alignment against a spatial reference. LUNA was released as a bioRxiv preprint in February 2025 and presented at the ICML 2025 Generative AI for Biology workshop.
LUNA is implemented in PyTorch and PyTorch Lightning. The denoising network is a stack of transformer layers with multi-head self-attention over the cells in a tissue graph, with separate input MLPs embedding node features, diffusion timestep, and current positions, and output MLPs producing the position update; torch-geometric supplies the graph handling. The training objective is a mean-squared error on pairwise Euclidean distance matrices rather than on absolute coordinates, so the model learns relative cell-cell geometry and its reconstructions are invariant to global rotation, translation, and the ordering of cells in the input.
Training and evaluation span several technologies: the Allen Brain Cell MERFISH atlas of the whole mouse brain (over 1.2 million cells), a mouse central nervous system scRNA-seq atlas reassembled de novo, and human melanoma Slide-tags data. Code, configuration files, preprocessed datasets, and a MERFISH mouse primary motor cortex sample are released under the MIT licence. No trained checkpoint ships with them: the repository's test-only mode evaluates a checkpoint the user has trained themselves, so applying LUNA to new data means running the training pipeline first.
The immediate use is retrospective: taking a dissociated atlas that a lab already owns and asking where those cells sat. That supports niche and cell-cell interaction analysis on data collected without a spatial assay, filling in nuclei dropped by barcode-based spatial protocols, and comparing tissue architecture between conditions when only expression is available for one arm. The Parkinson's disease result points to a diagnostic use as well — scoring how far a sample's inferred organisation deviates from a healthy reference to localise regions of change.
LUNA reframes spatial reconstruction as generative modelling of tissue geometry rather than as a mapping problem against a reference, and demonstrates that a single trained model transfers to cell types and disease states it never saw. The evidence base is currently mouse brain and one human melanoma sample, so performance in other organs and in tissues with less stereotyped architecture than the brain remains to be established, and the work is still a preprint. Code and preprocessed data are public under MIT, which lowers the barrier for the field to test those limits directly.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.