Lab Newsletter — August 13, 2026: The Cell's Neighborhood

Single-cell RNA sequencing changed biology by answering what is this cell? — one cell at a time, across the whole transcriptome. But to read a cell that way, you first have to dissolve the tissue into a suspension, and the moment you do, you lose the one thing a microscope never forgets: where the cell was, and who it was next to. A tumor is not a bag of cells; it is a neighborhood — immune cells pressing against a malignant core, a boundary where the fight actually happens. Dissociate it and that geography is gone. The frontier that gets it back reframes the question from what is this cell? to what is this cell, and where does it sit?

🗺️ Keep the address: reading RNA in place

The measurement came first. Imaging-based spatial transcriptomics — MERFISH (Chen, Zhuang et al., Science, 2015) and its relatives (seqFISH, in-situ sequencing, and the commercial Xenium and CosMx platforms) — images thousands of RNA species in single cells using combinatorial, error-robust fluorescence barcodes read out over many rounds of imaging, placing individual transcripts on a map of the tissue at subcellular resolution. Bulk sequencing gives an average; single-cell sequencing gives the parts list; spatial imaging gives the parts list with every part still pinned to its place. The catch is that it stops being a sequencing problem and becomes a microscopy one. As the field puts it plainly, “inaccurate cell segmentation procedures lead to misassignment of mRNAs to individual cells, which can introduce errors in downstream analysis.” You are looking at a dense scatter of glowing dots and you must decide, for each one, which cell does this belong to? — draw the boundaries wrong and every downstream cell type is wrong. Baysor (Nature Biotechnology) is a leading answer: it optimizes 2D or 3D cell boundaries by weighing the joint likelihood of transcriptional composition and cell morphology, working from RNA alone or with image-based priors — nuclei segmented by Cellpose, a membrane stain as a guide — to recover cells that boundary-only methods miss. Why it matters for the lab: transcript assignment is cell segmentation — the exact BioImage Model Zoo / Cellpose / SAM problem the lab works on, now load-bearing for omics.

🧬 Make sense of the map: spatial foundation models

Segmenting the cells only draws the map; the harder question is what the arrangement means. Nicheformer (Schaar, Tejada-Lapuerta, … Fabian Theis et al., Nature Methods, 2025) is a foundation model for single-cell and spatial omics — a transformer trained with a masked-language-modeling objective to learn a cell representation that captures spatial context: the niche, the neighborhood, who a cell sits among. It was pretrained on SpatialCorpus-110M, a corpus of over 110 million cells across 73 organs and tissues from human and mouse, deliberately pairing dissociated single-cell data with ~54 million cells measured by image-based spatial technologies. That pairing is the point, and it yields the paper’s sharpest finding: models trained only on dissociated data fail to recover the complexity of spatial microenvironments — location is not a decoration you can reconstruct after the fact, it is information you have to learn from spatial data directly. Nicheformer predicts spatial labels, niche identity, and local cell-density and composition, and can even transfer spatial context back onto plain scRNA-seq, hinting at where a dissociated cell would have lived. The code and pretrained weights are open (Hugging Face theislab/Nicheformer). Why it matters for the lab: open, runnable models spanning the imaging and omics halves of a spatial pipeline is the BioEngine ethos exactly.

🧭 The tissue under the cell — and the honest frontier

Here is why this belongs on a lab working toward a virtual cell. A Human Cell Simulator that treats each cell as an independent sample misses the central truth Nicheformer quantifies: a cell’s state depends on its neighbors. If Aug 12 added the molecular interactions between parts and Aug 11 added their motion, this adds the level above the cell — tissue context and niche, where disease actually unfolds. And it is, at bottom, a microscopy acquisition: imaging-based spatial transcriptomics is exactly the kind of multi-round imaging a self-driving microscope, Agent-Lens, or REEF imaging farm is built to run — and the native-tissue complement to Aug 10’s in-situ perturbation readout (that was a screen; this is a map). But the frontier stays honest, and that is what keeps it useful. Imaging-based panels are targeted — hundreds to a few thousand genes, not the whole transcriptome — so you see the questions you thought to ask. Segmentation errors propagate: a misdrawn boundary sends a transcript to the wrong cell and quietly corrupts every cell type built on top of it. And a foundation model’s predicted spatial context is a hypothesis, not a measurement — a claim to validate, not to trust on sight. That is the same prove-it discipline we keep returning to: a map is only as good as the boundaries you draw on it. Sequencing told us what a cell is. Spatial biology is starting to tell us what it is in context — who it lives beside, and how that changes what it does.

Sources linked inline. Compiled by Happy Agent; the lab footer notes our AI-assisted content. (X/Twitter sweep was skipped today — our news API is out of credits; a Grok-based replacement is wired and awaiting credits.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.