Lab Newsletter — July 26, 2026: Cells in Context

To sequence a cell you usually have to rip it out of its tissue — and the moment you do, you lose where it was and who its neighbors were. That spatial context is often the whole story. Today’s three items are all about giving it back.

🧫 A cheap slide, read for its molecules

The standout is Path2Space, published in Cell: a model that predicts the spatial expression of thousands of genes directly from an ordinary H&E histology slide — the same stained tissue image a pathologist has looked at for a century. Trained on breast-cancer spatial-transcriptomics data, it outperforms 21 established methods, and applied to 976 TCGA tumors it maps the tumor microenvironment, finds three new prognostic subgroups (“SpatioTypes”), and — most usefully — predicts response to chemotherapy and trastuzumab better than costly bulk-sequencing biomarkers, keying on how HER2 expression is spatially scattered across the tumor. Why it matters for the lab: this is our thesis in one result — a cheap image, read by AI for the expensive molecular measurement underneath. It’s exactly the image-to-molecule move behind our ProtiCelli work, and the recurring win of a good surrogate: skip the assay, keep the answer.

🧩 Putting dissociated cells back in place

The complement to predicting spatial data is recovering it. Nicheformer (Helmholtz Munich / TUM, Nature Methods) is a foundation model trained on over 110 million cells — a curated SpatialCorpus-110M blending 57M dissociated cells with 53M spatially resolved ones — that learns to transfer spatial context back onto single-cell data that never had any. Its quiet but important finding: spatial patterns leave measurable traces in gene expression even after cells are dissociated, so the neighborhood a cell came from can be partly reconstructed. Why it matters for the lab: it’s another argument that the payoff comes from model-ready, well-curated data — the same lesson the Virtual Cell Challenge drove home — and a reminder that the decades of dissociated single-cell atlases aren’t spatially blind after all.

🤖 An agent that runs the whole spatial workflow — and the honest asterisk

Then there’s automation. SpatialAgent (Genentech / Stanford) is an autonomous agent for spatial biology that pairs a language model with tool execution to run the loop end-to-end — design a gene panel, annotate cells and niches, generate hypotheses — in either fully autonomous or co-pilot mode. On ~2M cells it reportedly matched expert accuracy in heart-tissue annotation while cutting the time ~80%, and surfaced novel TGF-β fibroblast–pericyte interactions in colitis that earlier studies missed; a 2026 feature places it in a fast-commercializing field (GSK paid $50M to license spatial-cancer models this year). The honest asterisk: the headline claim that it “matched or outperformed human scientists” comes from a preprint, not peer-reviewed work. Why it matters for the lab: it’s the shape of things we build — agents that reason and act over microscopy and omics, like Agent-Lens and the BioImage.IO chatbot — and the caveat is the point: the agent proposes, but a REEF-style closed loop still has to check.

Predict the context from a cheap image, restore it onto old data, and let an agent work the whole board — but keep a human, and a wet lab, in the loop to say whether it’s right. Spatial biology is where imaging, omics and agents finally meet, which makes it about as on-brand as a week gets.

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.) Have lab news to share — a talk, paper, conference or release? Message me on Slack.