Lab Newsletter — September 3, 2026: The Cell in Every Language

A cell is not a single measurement. In the same instant it is doing several things at once: its chromatin is deciding which genes are even legible, its transcriptome is the set of RNAs it’s actively reading out, and its proteins are the molecules physically doing the work. Each is a different language describing the same cell — and for most of single-cell biology’s history we could only listen to one at a time, usually RNA. That has changed. Assays now read two or three layers in the same cell, and with that comes a genuinely new computational problem — not measuring a modality, but fusing them. As a 2021 review by Argelaguet and colleagues (Nature Biotechnology) frames it, “the development of single-cell multimodal assays provides a powerful tool for investigating multiple dimensions of cellular heterogeneity,” and “a key challenge in the analysis of single-cell multimodal data is to devise appropriate strategies for tying together data across different modalities.” Their honest conclusion — that “new definitions and concepts are needed to contextualize existing methods and to enable development of new methods” — is why this corner of the field has produced not one method but a small ecosystem.

🧬 Paired layers, fused per cell

Start with the friendliest case: two modalities measured in the same cell, like RNA and surface protein via CITE-seq. Even here, naïvely concatenating the numbers fails — the modalities have different scales, different noise, different reliability. Seurat’s weighted-nearest-neighbor (WNN) (Hao … Satija, Cell, 2021) answers with a beautifully simple idea: “an unsupervised framework to learn the relative utility of each data type in each cell.” Some cells are best distinguished by their proteins, others by their transcripts — so let the weighting be per cell. The authors applied it “to a CITE-seq dataset of 211,000 human peripheral blood mononuclear cells (PBMCs) with panels extending to 228 antibodies to construct a multimodal reference atlas of the circulating immune system,” and framed the payoff as exactly the shift this digest is about: looking “beyond the transcriptome toward a unified and multimodal definition of cellular identity.”

Where WNN combines views geometrically, totalVI (Gayoso … Yosef, Nature Methods, 2021) models them generatively. The trouble it names is the crux of the whole field: “combining these paired views into a unified representation of cell state is made challenging by the unique technical characteristics of each measurement.” totalVI’s answer is to represent the data “probabilistically … as a composite of biological and technical factors, including protein background and batch effects” — so the model can subtract the artifacts (antibody background, batch) and keep the biology. That separation of signal from instrument is the recurring theme in every good integration method.

🔓 More layers — and the ones you didn’t measure

RNA-plus-protein is only the start. MultiVI (Ashuach … Yosef, Nature Methods, 2023) brings in chromatin: it targets the joint profiling of “the transcriptome, chromatin accessibility and other molecular properties of single cells,” a probabilistic model whose most useful trick is what it does with incomplete data. MultiVI “creates a joint representation that allows an analysis of all modalities included in the multiomic input data, even for cells for which one or more modalities are missing.” That last clause matters more than it sounds: most existing datasets measured only one modality, so a model that can place single-modality cells into the same shared space as multiomic ones lets you reuse the enormous archive of older RNA-only or ATAC-only experiments instead of throwing it away.

🌉 Across datasets, with no shared cells at all

The hardest version drops the last comfort — the modalities aren’t even measured in the same cells. GLUE (Cao & Gao, Nature Biotechnology, 2022, open access) states the problem plainly: “despite the emergence of experimental methods for simultaneous measurement of multiple omics modalities in single cells, most single-cell datasets include only one modality,” and “a major obstacle in integrating omics data from multiple modalities is that different omics layers typically have distinct feature spaces.” Genes, peaks, and proteins simply don’t live in the same coordinate system. GLUE — “graph-linked unified embedding” — “bridges the gap by modeling regulatory interactions across omics layers explicitly,” using prior biological knowledge (which regions regulate which genes) as the bridge. Benchmarking showed it “more accurate, robust and scalable than state-of-the-art tools,” and the authors pushed it to “triple-omics integration, integrative regulatory inference and multi-omics human cell atlas construction over millions of cells, where GLUE was able to correct previous annotations.” Fusion at atlas scale, good enough to fix mistakes in the reference.

🧭 The honest frontier — integration you can trust

With this many methods, the field needed a referee, and it built a rigorous one. scIB (Luecken … Theis, Nature Methods, 2022) “benchmarked 68 method and preprocessing combinations on 85 batches of gene expression, chromatin accessibility and simulation data from 23 publications, altogether representing >1.2 million cells distributed in 13 atlas-level integration tasks.” Its evaluation crystallizes the central danger of all integration: you can always make datasets overlap by erasing what makes them different — so scIB scores methods on “their ability to remove batch effects while retaining biological variation,” across 14 metrics. The verdict is refreshingly specific ("scANVI, Scanorama, scVI and scGen perform well, particularly on complex integration tasks, while single-cell ATAC-sequencing integration performance is strongly affected by choice of feature space"). It’s the same prove-it discipline we keep returning to — the CAFA benchmark for protein function, the Ma benchmark for tissue segmentation, CryoBench: the score that counts measures whether you kept the biology, not just whether the plot looks clean.

And it lands squarely where we work. A serious virtual cell or Human Cell Simulator can’t be built on the transcriptome alone — it needs the multi-layer definition of a cell, chromatin and RNA and protein tied together, which is precisely what these methods learn. It’s also the molecular twin of the lab’s imaging×omics core: the spatial RNA map and protein map we told earlier are simply more modalities to fuse, in space as well as in molecule. And the way these tools travel is the ethos we keep pointing at — totalVI and MultiVI ship inside the open scvi-tools stack, GLUE is open on GitHub, and scIB is the public yardstick: the same publish-the-model-and-the-test-that-could-embarrass-it spirit behind the BioImage Model Zoo and BioEngine. Listen to a cell in every language it speaks, and teach a machine to hear them as one voice — that’s the representation the rest of the pipeline has been waiting for.

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