Lab Newsletter — September 23, 2026: Teaching Machines to Name Cells

We’ve spent the week among molecules — structures, immune receptors, drug responses. Today we step back up to the level of the whole cell, and a question so basic it’s easy to overlook: what kind of cell is this? Every single-cell experiment produces thousands of expression profiles, and for years a human squinted at clusters and hand-labeled them — “this looks like a T cell, that one’s a macrophage.” Today’s digest is about the quiet revolution that turned that manual art into an automated, reference-driven science, and why it’s the labeling layer beneath the virtual cell: you cannot model a cell you cannot first name.

🗺️ The vision: a complete catalogue of human cells

It starts with an audacious goal. Regev, Teichmann et al. (eLife, 2017) launched the Human Cell Atlas with the conviction that “the recent advent of methods for high-throughput single-cell molecular profiling has catalyzed a growing sense in the scientific community that the time is ripe to complete the 150-year-old effort to identify all cell types in the human body.” The plan: “to define all human cell types in terms of distinctive molecular profiles” and connect them to “classical cellular descriptions (such as location and morphology).” Crucially, they built it around “a commitment to open data, code, and community” — the same open-infrastructure creed behind the lab’s BioImage Model Zoo and BioEngine, now aimed at cells.

📖 The reference: half a million cells, 400+ types

A vision needs a concrete reference to point to. The Tabula Sapiens Consortium (Science, 2022) delivered one, noting that “molecular characterization of cell types using single-cell transcriptome sequencing is revolutionizing cell biology.” They “created a human reference atlas comprising nearly 500,000 cells from 24 different tissues and organs, many from the same donor,” which “enabled molecular characterization of more than 400 cell types, their distribution across tissues, and tissue-specific variation in gene expression.” This is the object every annotation method leans on: a shared, curated map of what cells are, against which any new dataset can be read.

🏷️ The first automation: annotate by reference

With references in hand, why label by hand at all? Aran et al. (Nature Immunology, 2019) introduced SingleR as exactly that shortcut — “a novel computational framework for the annotation of scRNA-seq by reference to bulk transcriptomes.” And it wasn’t just bookkeeping: automated typing let them subcluster macrophages and reveal “a disease-associated subgroup with a transitional gene expression profile intermediate between monocyte-derived and alveolar macrophages,” cells that “localized to the fibrotic niche and had a profibrotic effect in vivo.” A lesson that recurs across the lab’s work: good automation doesn’t just save time — it finds things humans miss.

⚡ Scaling it up: a machine-learning cell-typer

As datasets grew to hundreds of thousands of cells across many tissues, annotation itself had to become a fast, learned model. Domínguez Conde, Teichmann et al. (Science, 2022) built CellTypist for that — “a machine learning tool for rapid and precise cell type annotation” — to “systematically resolve immune cell heterogeneity across tissues” over “a dataset of ~360,000 cells.” Their approach “lays the foundation for identifying highly resolved immune cell types by leveraging a common reference dataset, tissue-integrated expression analysis, and antigen receptor sequencing” — annotation as a reusable, pretrained classifier rather than a bespoke analysis each time.

🎲 The probabilistic turn: labels with uncertainty

Cells are noisy, and a confident wrong label is dangerous. Xu, Lopez et al. (Molecular Systems Biology, 2021) brought deep generative modeling to the problem with scANVI, framing the goal plainly: as datasets accumulate, “the natural next step is to integrate the accumulating data to achieve a common ontology of cell types and states,” yet “it is not straightforward … to automatically assign cell type labels in a new dataset based on existing annotations.” Their answer is “single-cell ANnotation using Variational Inference (scANVI), a semi-supervised variant of scVI designed to leverage existing cell state annotations” — one that models “uncertainty caused by biological and measurement noise.” It’s the same representation-learning bet the lab makes across proteins and imaging, now producing a probabilistic label, not a guess.

🔐 Map, don’t move: references without sharing raw data

The last step is the one that makes atlases a shared resource: how do you annotate your cells against a reference you don’t own, without shipping sensitive raw data around? Lotfollahi et al. (Nature Biotechnology, 2022) answered with scArches — “single-cell architectural surgery” — a transfer- learning strategy that maps “query datasets on top of a reference … without sharing raw data.” It “preserves biological state information while removing batch effects, despite using four orders of magnitude fewer parameters than de novo integration,” and strikingly, it “retains coronavirus disease 2019 (COVID-19) disease variation when mapping to a healthy reference, enabling the discovery of disease-specific cell states.” Decentralized, privacy-aware reference building — precisely the principle behind the lab’s Safe Colab and BioEngine.

🧫 Why it’s our kind of problem

Read across the six and a pattern emerges that is squarely the lab’s own. First, a cell atlas is the data foundation of the virtual cell: before you can simulate how a cell changes state — under a drug, a signal, a mutation — you need a shared, machine-readable ontology of what states exist. Annotation builds that vocabulary. Second, the winning designs are open references plus reusable models (HCA’s open-data commitment, scvi-tools, CellTypist as a pretrained classifier) — the same bet the lab makes with its model zoo and shared infrastructure. Third, scArches’ “map without sharing raw data” is the federated, privacy-preserving principle the lab is building into Safe Colab, so that hospitals and labs can contribute to a common reference without surrendering their data. Naming cells sounds mundane. It is, in fact, the first sentence in the language we’ll need to write down a cell.

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