CytoVI: Deep learning model unifies antibody-based single-cell data across platforms

Antibody-based single-cell technologies have transformed how researchers and clinicians inspect the immune system, yet the data they produce have long resisted unified analysis. Flow cytometry, mass cytometry and CITE-seq each measure proteins on millions of individual cells, but they do so with different reporters, different panels and different sources of technical noise. A team led by Florian Ingelfinger, Ido Amit and Nir Yosef, working across the Weizmann Institute of Science, UC Berkeley and partner clinical centers, now reports in Nature Methods a deep generative model called CytoVI that promises to knit these fragmented datasets into a single, statistically coherent picture. The software is released as open source within the widely used scvi-tools framework.

The core problem CytoVI addresses is that antibody-based measurements are relative, not absolute. In flow cytometry, a protein’s signal depends on which fluorophore is attached to the antibody; in mass cytometry, on which heavy-metal isotope is used. Because laboratories choose different reporter molecules, staining protocols and antibody panels, the same cell type can look strikingly different across experiments. Batch effects arising from instrument performance, reagent lots and sample handling further obscure biology. Comparative studies across the vast public repositories of cytometry data have therefore been largely impractical, since different studies interrogate different marker sets with different technologies.

CytoVI is a variational autoencoder, a probabilistic latent variable model in which an encoder neural network compresses each cell’s protein expression profile into a low-dimensional latent vector that represents the cell’s intrinsic state, complete with an associated uncertainty. A decoder network then reconstructs the original protein expression from that embedding, conditioned on observed nuisance covariates such as batch, technology or antibody panel. By optimizing the evidence lower bound, a combination of a reconstruction term and a Kullback–Leibler regularization, the model learns embeddings that are corrected for technical variation while retaining biological signal. Depending on the expected heterogeneity of the sample, the latent space is assigned either an isotropic Gaussian prior or a more expressive mixture of Gaussians, which can optionally be informed by existing cell type labels in difficult integration scenarios.

Because the model is generative, it does far more than integrate. The trained decoder can produce batch-corrected expression estimates, test for differentially expressed proteins within a Bayesian framework, and, crucially, impute proteins that were never measured. When two experiments use partially overlapping antibody panels, the encoder reads only the shared backbone markers while the decoder reconstructs the union of all markers, effectively filling in the gaps with probabilistic estimates that carry explicit uncertainty. The same counterfactual decoding logic allows CytoVI to impute morphological features such as forward and side scatter into mass cytometry data, and even to impute transcriptomes into flow cytometry datasets when paired CITE-seq measurements are available as a reference.

The authors validated the model through a series of rigorous benchmarks. Posterior predictive checks showed that CytoVI faithfully reproduces the statistical properties of flow cytometry, mass cytometry and CITE-seq data. In integration tests on replicate peripheral blood mononuclear cell samples measured in different batches, CytoVI removed technical variation while preserving cell type structure, performing at least on par with state-of-the-art tools such as cyCombine, FastMNN and Harmony, and avoiding the residual batch effects those methods left in specific immune compartments. The model remained robust even when more than two-thirds of the markers in one batch were masked, and training stayed computationally manageable, taking under 35 minutes for one million cells on a single GPU.

To demonstrate the scale of insight the approach enables, the team assembled an integrated B cell maturation atlas from mass cytometry data spanning twelve partially overlapping antibody panels. By controlling for batch and imputing non-overlapping markers, CytoVI characterized 350 surface proteins across transitional, naive and memory B cell compartments. Trajectory modeling with diffusion maps and CellRank revealed the differentiation path from transitional B cells toward class-switched memory cells, and the model’s differential expression module pinpointed proteins associated with immunoglobulin class-switching, including CD62L, CD71 and CD100, molecules known to mediate lymph node migration, antigen specificity and B cell–T cell communication.

The clinical potential of the method became apparent in analyses of patient cohorts. In a study of lymph node T cell infiltrates from 63 patients with B cell non-Hodgkin lymphoma, profiled with two overlapping flow cytometry panels, CytoVI’s label-free differential-abundance analysis automatically identified disease-associated T cell states. It reproduced known findings, such as follicular helper T cell enrichment in follicular lymphoma, but also resolved a proliferating population as predominantly CD8-positive T cells and uncovered a small cluster of CD4/CD8 double-negative T cells enriched in diffuse large B cell lymphoma and mantle cell lymphoma that the original manual analysis had missed. By integrating the flow data with paired CITE-seq measurements, the team imputed transcriptomes and showed that these double-negative cells are in fact cytotoxic Vδ1 gamma-delta T cells, a population previously reported to be enriched in lymphoma patients.

Perhaps most strikingly, CytoVI was applied to routine diagnostic flow cytometry for chronic lymphocytic leukemia. Training a reference model on a small set of patients and controls, the researchers used transfer learning to map new, unseen diagnostic samples into the reference latent space without retraining. Automatically detected CLL cell frequencies correlated strongly with expert hematologist annotations, and the kappa-to-lambda light-chain ratio computed within the automatically classified cells captured the monoclonality that is a diagnostic hallmark of the disease. Combining these two features in a simple classifier distinguished CLL patients, including those with minimal residual disease, from controls with an area under the curve of about 0.74, a task that remains challenging even for experts.

The authors frame CytoVI as a step toward a model-centric view of cytometry, analogous to the shift already underway in single-cell transcriptomics. Because cytometry is destructive and restricted to predefined panels, proteins absent from an original experiment are normally lost forever; CytoVI’s probabilistic imputation breathes new life into archived datasets and could support community-scale protein atlases built from conventional assays. With intrinsic uncertainty estimates suited to privacy-sensitive clinical deployment, interoperability with Scanpy workflows, and the ability to fuse the scale of cytometry with the molecular resolution of single-cell genomics, the model lays groundwork for AI-powered cytometry that could accelerate biomarker discovery, standardize diagnostic interpretation and systematically chart cellular states across health and disease.

Subject of Research: Deep generative modeling for unified analysis of antibody-based single-cell data

Article Title: CytoVI: deep generative modeling of antibody-based single cell data

Article References: Ingelfinger, F., Levy, N., Ergen, C., Bakulin, A., Becker, A., Boyeau, P., Kim, M., Ditz, D., Dirks, J., Maaskola, J., Wertheimer, T., Zeiser, R., Widmer, C. C., Amit, I., & Yosef, N. (2026). CytoVI: deep generative modeling of antibody-based single cell data. Nature Methods, 23(10), 2054-2065. https://doi.org/10.1038/s41592-026-03224-5

Image Credits: AI Generated

DOI: 10.1038/s41592-026-03224-5

Keywords: CytoVI, flow cytometry, mass cytometry, CITE-seq, deep generative model, variational autoencoder, batch correction, protein imputation, B cell atlas, non-Hodgkin lymphoma, chronic lymphocytic leukemia, scvi-tools