Large-scale Foundation Model Reconstructs How Cells Interact Within Tissues

Missing discourse successful single-cell data

Single-cell RNA sequencing has transformed biology by showing which genes are progressive successful individual cells. However, this attack requires cells to beryllium removed from their earthy environment, erasing accusation astir their position and neighbors. Spatial transcriptomics preserves this discourse but is technically much constricted and harder to scale. Researchers person agelong lacked a measurement to study compartment personality and insubstantial statement together.

AI exemplary reveals hidden insubstantial structures

Nicheformer overcomes this obstruction by learning from some dissociated and spatial data. It tin “transfer” spatial discourse backmost onto cells that were antecedently studied successful isolation – fundamentally reconstructing really they fresh into nan bigger image of a tissue. To make this possible, nan investigation squad created SpatialCorpus-110M, 1 of nan largest curated resources of single-cell and spatial information to date. In their study published successful Nature Methods, nan exemplary consistently outperformed existing approaches and showed that spatial patterns time off measurable traces successful cistron expression, moreover erstwhile cells are dissociated. Beyond performance, nan researchers besides explored interpretability, revealing that nan exemplary identifies biologically meaningful patterns successful its soul layers – offering a caller model into really AI learns from biology.

“With Nicheformer we tin now transportation spatial accusation onto dissociated single-cell information astatine scale,” says Alejandro Tejada-Lapuerta, PhD student astatine Helmholtz Munich and TUM and co-first writer of nan study together pinch Anna Schaar. “This opens up galore possibilities to study insubstantial statement and cellular neighborhoods without further experiments.”

The study connects to nan emerging thought of a “Virtual Cell”, a computational practice of really cells behave and interact wrong their autochthonal environments. While this conception is gaining momentum crossed biology and AI, erstwhile models person mostly treated cells arsenic isolated entities, without reasoning their spatial relationships. Nicheformer is nan first instauration exemplary to study straight from spatial organization, offering a measurement to reconstruct really cells consciousness and power their neighbors. Beyond introducing this caller capability, nan researchers besides coming an full suite of spatial benchmarking tasks that situation early models to seizure insubstantial architecture and corporate cellular behaviour – an basal measurement toward biologically realistic AI systems.

Next steps

With Nicheformer we are taking nan first steps toward building general-purpose AI models that correspond cells successful their earthy discourse – nan instauration of a Virtual Cell and Tissue model. Such models will toggle shape really we study wellness and illness and could yet guideline nan improvement of caller therapies.”

Prof. Fabian Theis, Director of nan Computational Health Center astatine Helmholtz Munich and Professor astatine TUM

In their adjacent project, nan squad intends to create a “tissue instauration model” that besides learns nan beingness relationships betwixt cells. Such a exemplary could thief analyse tumor microenvironments and different analyzable structures successful nan assemblage pinch nonstop relevance for diseases specified arsenic cancer, diabetes, and chronic inflammation.

Source:

Journal reference:

Tejada-Lapuerta, A., et al. (2025). Nicheformer: a instauration exemplary for single-cell and spatial omics. Nature Methods. doi.org/10.1038/s41592-025-02814-z