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

Abstract

Antibody-based single-cell technologies, such as flow cytometry, mass cytometry and CITE-seq, have become widely used in clinical diagnostics and basic research; however, their analysis is complicated by technical noise, batch effects, platform differences and restricted antibody panels. Here we present CytoVI, a probabilistic generative model for statistically rigorous unified analysis of antibody-based single-cell data. CytoVI generates informative cell embeddings, imputes missing measurements, performs differential protein expression testing and automates annotation of cells in a single probabilistic model. We applied CytoVI to build an integrated B cell maturation atlas spanning 350 proteins and identified proteins associated with immunoglobulin class-switching. In a cohort of patients with B cell non-Hodgkin lymphoma profiled by flow cytometry and CITE-seq, CytoVI uncovered disease-associated T cell states. CytoVI is available as open-source software at scvi-tools.org.

CytoVI is a deep generative model for statistically rigorous analysis of antibody-based single-cell data.

Main

Single-cell multi-omics have been applied extensively to assemble comprehensive molecular atlases of individual cells within our body¹ and to unveil molecular and cellular perturbations that may lead to the manifestation of disease or explain a patient’s response to therapy². Flow cytometry was historically the first technology capable of analyzing millions of individual cells within minutes³. Due to its low cost, dynamic range, robustness and scalability, flow cytometry has become a cornerstone of clinical practice, facilitating the diagnosis of hematological malignancies or serving as a prognostic tool to guide treatment decisions.

Since its introduction in 1965, flow cytometry has evolved into a powerful and cost-effective tool for exploratory research, now enabling the simultaneous measurement of more than 50 cellular features⁴. Yet further increasing its dimensionality remains technically challenging because of fluorophore spectral overlap. Alternative technologies such as mass cytometry (CyTOF) and CITE-seq have expanded protein profiling through heavy metal- or oligonucleotide-tagged antibodies^5,6, but their broader application has been limited by labeling constraints and cost.

The combination of machine-learning and traditional flow cytometry demonstrated the tremendous potential to computationally mitigate the limited number of cellular features detectable, but these approaches were not readily applicable to multi-donor settings due to the requirement of a high number of cells per assay⁷. Scyan, a Bayesian model that incorporates biological priors on marker positivities of expected cell types, highlighted the utility of deep probabilistic modeling for the automated annotation of cytometry data⁸. More broadly, deep generative latent variable models have emerged as powerful tools for the integration, denoising and imputation of single-cell genomics data^9,10. In contrast to the stochastic nature and sparsity associated with genomic single-cell technologies, antibody-based cytometry technologies offer high-resolution measurements of millions of cells across a wide dynamic range, facilitating the analysis of large patient cohorts; however, the analysis of multi-cohort studies is often obstructed by batch effects attributed to differences in sample isolation protocols or antibody staining conditions. Moreover, comparative studies using the vast amount of publicly available cytometry data are hampered by the use of different technologies and distinct antibody panels to interrogate samples.

Here, we present CytoVI, a probabilistic end-to-end deep learning framework designed for the integration of antibody-based single-cell technologies. CytoVI removes technical variation in flow cytometry, mass cytometry or CITE-seq data and embeds cells into a meaningful low-dimensional representation. We demonstrate that CytoVI is capable of imputing missing markers in experiments with overlapping antibody panels, imputing a cell’s transcriptome when paired with CITE-seq data, and facilitating the automated detection of cellular states associated with clinical covariates, such as a patient’s response to treatment. CytoVI is available as an open-source Python library at scvi-tools.org (ref. ¹¹), leveraging advanced optimization strategies for the integration of large single-cell datasets. CytoVI overcomes major limitations of modern cytometry and lays the groundwork for next-generation AI-powered cytometry integrated with other single-cell genomic modalities, with broad potential for biomarker discovery and clinical utility.

Results

The CytoVI model

Flow cytometry is the most abundant single-cell modality in the public domain and has great potential for cellular atlassing of thousands of individuals across billions of cells; however, efficient large-scale integration of antibody-based single-cell technologies is hampered by technical variation introduced during data generation. To decouple technical variation from biological variation in a unified manner for antibody-based single-cell technologies, CytoVI encodes each cell’s protein expression in a shared normally distributed latent space that corresponds to its intrinsic state with its associated uncertainty (Fig. 1a). In a subsequent step, CytoVI generates probabilistic protein expression estimates for each cell, given its embedding coordinates. Additionally, CytoVI can incorporate observed nuisance factors that inform technical variation due to experimental batches, differences in the applied technology, or confounding biological variables such as sex or age. These artifacts are controlled for in the inferred embedding and in the generative process, thus giving rise to corrected cell state and protein expression estimates. CytoVI can also incorporate a priori cell type labels, which are used to inform the prior distribution of the embedding space (Methods), a step that we found helpful in difficult integration scenarios¹².

Fig. 1. Schematic overview of CytoVI.

a, CytoVI is a latent variable model consisting of an encoder and a decoder neural network. The encoder maps the protein expression matrix of antibody-based single-cell technologies (X) into a joint latent space (z). Depending on the expected underlying biology of the data we assume an isotropic Gaussian prior or a mixture of Gaussians prior for the posterior distribution p(z∣x). The decoder network incorporates the posterior mean of (z) and observed sample-level covariates such as batch (s) and fits the parameters of a Beta or Gaussian distribution for the protein likelihood to reconstruct the original protein expression. The variational and neural network parameters are trained by maximizing the evidence lower bound (ELBO) consisting of a reconstruction term and Kullback–Leibler (KL) divergence. Thereby, CytoVI optimizes a latent representation for each cell that is corrected for the observed sample-level covariates (s). b, Schematic overview of the analysis tasks that can be accomplished using CytoVI. Panels created in BioRender: a, Ingelfinger, F. https://biorender.com/89fypas (2026); b, Ingelfinger, F. https://biorender.com/3bsx4fm (2026).

Both the encoder, which facilitates inference of latent distributions, and the decoder, corresponding to the generative model that estimates the distribution of a cell’s protein expression, are composed of neural networks. Amortized variational inference estimates the neural network parameters and the variational parameters of the latent space distribution and protein expression likelihood. After training, the CytoVI latent space can be utilized for data integration, cell type assignment or label-free differential-abundance analysis. Conversely, the generative model can be utilized to obtain corrected expression estimates, identify differentially expressed proteins or impute missing markers when overlapping antibody panels are applied (Fig. 1b). In the latter setting, the inference network uses only the shared overlapping markers to optimize a cell’s latent distribution, whereas the generative network reconstructs the union of all markers present across the antibody panels, thereby effectively imputing unseen markers (Methods).

Antibody-based single-cell measurements capture relative signal intensities that depend on reporter molecules and therefore require transformation and scaling for downstream analysis. Accordingly, CytoVI models transformed protein expression using a Gaussian likelihood by default (Supplementary Note 1).

CytoVI accurately models antibody-based single-cell data and provides a biologically interpretable latent space

Posterior predictive checks demonstrated that CytoVI is capable of generating flow cytometry, mass cytometry and CITE-seq data that closely fit the observed data (Supplementary Note 2). Next, we evaluated the biological interpretability of CytoVI’s latent space and the extent to which it captures cell state variability across the different technologies. For this, we applied clustering to the protein expression (unintegrated data space) and used key protein markers to manually assign cell type labels for the human peripheral blood mononuclear cell (PBMC) datasets measured using flow cytometry, mass cytometry and CITE-seq^13–15 (Extended Data Fig. 1a,b). In cases where cell type labels had been provided in the original study, these annotations were retained (Extended Data Fig. 1c)¹⁵. Visualizing these cell type annotations in the CytoVI latent space confirmed its applicability for resolving cell type variation across flow cytometry, mass cytometry and CITE-seq data (Fig. 2a). To evaluate whether the CytoVI latent space could be employed for de novo cell type annotation, we performed Leiden clustering on the CytoVI latent space. Clusters obtained from the CytoVI latent space largely replicated the prior cell type annotations, including rare subsets such as dendritic cells (<2%), while avoiding over-fragmentation of larger populations as observed in other computational workflows for cytometry analysis (Fig. 2b and Extended Data Fig. 1d,e)^16,17. The CytoVI latent space also captured fine-grained cell state variation beyond manual annotations, exemplified in CD56^dim natural killer (NK) cells by a gradient in which higher CD16 expression paralleled TIM-3 and opposed Granzyme B (Fig. 2c), consistent with previous reports of TIM-3 upregulation following Fc-receptor engagement of CD16 and reduced cytotoxicity^18,19. Together, we demonstrate that CytoVI optimizes a biologically interpretable latent space, capturing a cell-intrinsic state that can be utilized for cell type annotation or for the identification of gradual protein expression changes.

Extended Data Fig. 1. CytoVI controls for technical variation but preserves biological variation (integration benchmarking).

A, B, C: UMAP (left panel) of the data space of PBMCs measured by flow cytometry (A, Nunez et al.), mass cytometry (B, Ingelfinger et al.), and CITE-seq (C, Hao et al.). Cells are colored by the ground truth cell type annotations that have been either provided with the original publication of the data (C) or were manually annotated (A, B). Heatmap (right panel) displaying the column-scaled mean protein expression intensities for the indicated immune populations. D: UMAP of the CytoVI latent space for PBMCs measured by flow cytometry (left panel). Cells are colored by the Leiden clusters obtained from the CytoVI latent space. Heatmap (right panel) displaying the column-scaled mean protein expression intensities for the indicated immune populations obtained from clustering the CytoVI latent space. E: UMAP of the CytoVI latent space for PBMCs measured by flow cytometry colored by their ground truth cell type label. F: UMAP of the unintegrated data space, the cyCombine-corrected expression, and the integrated embeddings obtained from Harmony, FastMNN, and CytoVI for two biological replicates of the flow cytometry samples measured in two different batches. Cells are colored by batch and the ground-truth cell-type annotation (displayed in A). In each of the three cases, the arcsinh-transformed data has been feature-wise z-score scaled prior to batch integration. G: Scatter-plot displaying the runtime in relation to the number of input cells for the downsampled flow cytometry dataset for each method. All methods were evaluated on a computation node equipped with 64 GB of system RAM, a 16-core CPU, and a single NVIDIA A100-SXM4 GPU with 80 GB of memory.

Fig. 2. CytoVI optimizes a biologically interpretable latent representation for antibody-based single-cell technologies that controls for technical variation (integration benchmarking).

a, Uniform Manifold Approximation and Projection (UMAP) of the CytoVI latent space for PBMCs measured by flow cytometry (left; Nuñez et al., 35 protein markers + FSC/SSC), mass cytometry (middle; Ingelfinger et al., 36 protein markers) and CITE-seq (right; Hao et al., 228 protein markers). Cells are colored by the ground-truth cell type annotations that have been either provided with the original publication of the data (right) or were manually annotated. b, Heatmap displaying the confusion matrix between the ground-truth cell-type annotations of the flow cytometry dataset and clusters obtained by clustering the CytoVI latent space. c, UMAP of the CytoVI latent space of the flow cytometry dataset indicating the expression for CD16, Tim-3 and Granzyme B. d, UMAP of the unintegrated data space of two biological replicates of flow cytometry samples measured in two different batches (Nuñez et al., 35 protein markers + FSC/SSC, identical panel), colored by batch and the respective cell type annotation (left). UMAP of the CytoVI latent space of the same data after controlling for the observed batch covariate (right). In both cases, the arcsinh-transformed data has been min–max scaled to ensure equal representation of each marker. e, Overview table benchmarking the integration performance of CytoVI, cyCombine, FastMNN and Harmony for the flow cytometry dataset (Nuñez et al., 35 protein markers + FSC/SSC, identical panel) in relation to different preprocessing choices. Different metrics are presented for bio-conservation and batch correction, and an aggregated consensus score summarizes the results. NMI, normalized mutual information; ARI, Adjusted Rand Index; cLISI, cell-type Local Inverse Simpson’s Index; iLISI, integration local inverse Simpson’s index; KBET, k-nearest-neighbor Batch Effect Test; PCR, principal component regression; T_H, helper T; T_reg, regulatory T; Mono, monocytes.

CytoVI removes technical variation while preserving biological variation

Batch effects, arising from differences in experimental protocols, staining conditions, instrument performance or reagent lots, are a major challenge in flow cytometry analyses²⁰. Therefore, we next benchmarked CytoVI’s integration performance. For this, we utilized replicates of cryopreserved PBMCs that had been analyzed in two distinct batches by flow cytometry and showed a clear batch effect¹³. When explicitly controlling for the observed batch covariate, CytoVI yielded a latent space that integrated both batches while preserving cell type variation (Fig. 2d). Next, we benchmarked CytoVI’s integration performance against cyCombine, a recently proposed method for the integration and imputation of cytometry data, and Harmony and FastMNN, two integration methods introduced for single-cell genomics that have previously been applied to cytometry data^21–23. We evaluated different metrics for batch correction and conservation of biological information for each integration methodology using the single-cell integration benchmark (scIB) framework²⁴. As some of these methods required specific scaling approaches for effective integration, we benchmarked each method across all available preprocessing options (Supplementary Note 3). CytoVI and FastMNN both led to a clear improvement across all three standardization approaches and generally compared favorably to the other two methods (Fig. 2e and Extended Data Fig. 1f). Yet, cyCombine and FastMNN could not account for technical variation in the myeloid compartment, whereas Harmony maintained an apparent batch effect in the T cell compartment, which was not observed with CytoVI (Extended Data Fig. 1f). As expected for a generative, probabilistic model, CytoVI incurred a higher training time than the comparator methods when used solely for integration, but remained computationally manageable (<35 min for 1 million cells) (Extended Data Fig. 1g). When comparing CytoVI to the cell-matching algorithm MARIO²⁵, we observed that MARIO integrated only a fraction of input cells and introduced a cell type bias in the integrated dataset (Supplementary Note 4).

Next, we assessed CytoVI’s sensitivity to hyperparameter choices and found that 20 latent dimensions and 128 hidden nodes gave the best balance between batch correction and biological conservation (Extended Data Fig. 2a). Lastly, to assess robustness to feature dropout, we progressively masked random markers in one batch and found that CytoVI maintained stable label preservation and batch correction until more than 29 of 41 markers were removed (Extended Data Fig. 2b).

Extended Data Fig. 2. Benchmarking of CytoVI and comparator methods and evaluation of sensitivity to model hyperparameters and backbone marker overlap (integration benchmarking).

A: Heatmaps displaying the impact of latent dimensionality and the number of hidden nodes in CytoVI’s integration performance for the Nunez et al. dataset. Integration was benchmarked using scIB, and the total scores (left panel), bio conservation scores (middle panel), and batch correction scores (right panel) are shown. B: Scatter plots showing the aggregated scIB metrics scores for bio conservation (left panel), batch correction (middle panel), and total score (right panel) in relation to a variable number of features that have been randomly masked in one of the batches from the Nunez et al. dataset before model training.

Together, our benchmarking of integration performance demonstrates that CytoVI excels at integration tasks and performs at least on par with current state-of-the-art integration tools for cytometry data.

CytoVI facilitates accurate imputation of unobserved proteins

CytoVI’s generative model comprises an intrinsic ability to generate unobserved data and thus to impute missing proteins from cytometry experiments in which different antibody panels have been applied (Methods). Imputation benchmarking showed that CytoVI achieves state-of-the-art imputation of unobserved proteins and provides probabilistic estimates of protein expression that capture imputation uncertainty (Supplementary Note 5).

To demonstrate CytoVI’s ability to integrate distinct experiments with partially overlapping antibody panels, we jointly analyzed two PBMC flow cytometry datasets from Nuñez et al.¹³ and Kreutmair et al.²⁶ and incorporated cell type labels to inform the latent space prior (Methods; Fig. 3a). CytoVI corrected for technical variation between studies while preserving donor-specific cell state variation, for instance in γδ T cells and plasmablasts, which can vary substantially between individuals (Fig. 3b and Extended Data Fig. 3a)^27,28. Imputation of non-overlapping markers increased the effective panel size and enabled deeper immunophenotyping, for example resolving CD56^dimCD16⁺ NK cell states defined by KLRG1, CD161 and CD57, the latter being expressed only by a small subpopulation (Fig. 3c and Extended Data Fig. 3b). This is consistent with a previous single-cell RNA-sequencing atlas describing CD57⁺ NK cells as a distinct mature NK cell subset in human peripheral blood²⁹. Merging of the two antibody panels using CytoVI accumulated knowledge about protein expression across both datasets, thereby enabling a more comprehensive immunophenotyping and illustrating how partially overlapping panels can be leveraged to assemble large-scale protein atlases. Of note, integration was also feasible without prior cell type information (Extended Data Fig. 3c). Leave-one-out ablation experiments showed that CytoVI imputes from a multi-marker latent structure rather than using individual proteins as shortcuts for imputation (Supplementary Note 6).

Fig. 3. CytoVI integrates cytometry data from distinct antibody panels and imputes missing proteins (imputation benchmarking).

a, Schematic overview for the integration of two distinct flow cytometry studies from Nuñez et al. (35 protein markers + FSC/SSC) and Kreutmair et al. (30 protein markers) with a shared overlap of 15 backbone protein markers. b, UMAP of the CytoVI latent space colored by the respective study. c, UMAP of the CytoVI latent space displaying the imputed expression of CD161, CD57 (both measured in the Kreutmair et al. study) and KLRG1 (measured in the Nuñez et al. study). d, UMAP of the integrated CytoVI latent space colored by harmonized cell type annotations between the two studies (left). Heatmap showing the imputed protein expression profile for the indicated cell types across the union of proteins measured in both studies. e, Schematic overview of the integrated B cell maturation atlas from mass cytometry data generated by Glass et al. (12 distinct antibody panels of approximately 39 protein markers with an overlap of 9 protein markers across all panels). f, Diffusion maps of the imputed protein expression profile of 350 surface markers showing the cell–cell transition matrix obtained from CellRank (left) and a manual summary of the B cell maturation trajectories (right). Coloring indicates cell type annotations obtained from clustering the imputed B cell maturation atlas. g, Heatmap displaying the imputed protein expression of the top-50 markers that correlate best with the differentiation trajectory from transitional B cells to class-switched memory B cells along the diffusion pseudotime. h, Line plots displaying the expression of the indicated immunoglobulins for the class-switched (teal) and non-class-switched memory B cell lineage (blue) along the diffusion pseudotime. i, Volcano plot displaying the differentially expressed proteins between class-switched (teal) and non-class-switched (blue) memory B cells determined using CytoVI’s differential protein expression module. j, Line plots displaying the expression of the indicated proteins for the class-switched (teal) and non-class-switched memory B cell lineage (blue) along the diffusion pseudotime. Panels created in BioRender: a, Ingelfinger, F. https://biorender.com/h16xxwo (2026); e, Ingelfinger, F. https://biorender.com/3aqgz3z (2026).

Extended Data Fig. 3. CytoVI integrates cytometry data from distinct antibody panels and imputes missing proteins (imputation benchmarking).

A, B: UMAP of the integrated CytoVI latent space of PBMCs from two distinct flow cytometry studies (Nunez et al., Kreutmair et al.) displaying the imputed expression of the indicated proteins. C: UMAP of the integrated CytoVI latent space of the distinct flow cytometry studies colored by study (left panel) and harmonized cell type labels (right panel). The CytoVI model has been trained without the use of a cell type-informed prior for comparison. D: Kernel density estimate plots displaying the uncorrected expression of the B cell markers in peripheral human B cells measured across 12 overlapping mass cytometry panels and two donors (resulting in 24 unique samples; Glass et al.). The displayed markers represent the backbone markers that are shared across the 12 antibody panels. E: UMAP of the CytoVI latent space of the B cell dataset colored by sample identifier. The color legend is displayed in D. F: UMAP of the CytoVI latent space of the B cell dataset displaying the batch-corrected expression of the backbone markers. G: Diffusion map (left panel) of the imputed protein expression profile of 350 surface markers of the B cell maturation atlas. Coloring indicates cell type annotations obtained from clustering the imputed data space followed by manual annotation. Heatmap (right panel) displaying the column-normalized mean imputed protein expression profiles of the 50 most variable markers between the indicated B cell subsets. Expression of CD22 has been manually added. Backbone markers that were measured in all of the antibody panels are highlighted in bold. H: Violin plots displaying the frequency of indicated B cell subsets for the two healthy donors measured across the different antibody panels (n = 12). P values correspond to a paired Wilcoxon test for the indicated B cell populations between the donors. I: Diffusion maps of the integrated B cell maturation atlas displaying diffusion pseudotime and Palantir pseudotime. J: Diffusion maps displaying the initial and terminal states of the B cell trajectory determined using CellRank.

In conclusion, our imputation benchmarking showed that CytoVI robustly integrates flow cytometry data across distinct studies and imputes unobserved proteins with accompanying uncertainty estimates, thereby expanding the protein space that can be queried in a joint downstream analysis (Fig. 3d).

CytoVI assembles an integrated B cell maturation atlas spanning 350 protein markers

Next, we evaluated the extent to which CytoVI could be leveraged to construct large-scale protein atlases from conventional cytometry assays. To this end, we analyzed mass cytometry data of circulating B cells, measured across 12 partially overlapping antibody panels (each targeting ~39 proteins; Fig. 3e)³⁰. When explicitly controlling for the batch covariate, CytoVI generated a latent space that minimized technical variation while preserving biological variation across B cell states (Extended Data Fig. 3d–f). Clustering of this latent space and imputation of the non-overlapping protein markers enabled the in-depth characterization of 350 surface proteins across transitional, naive, non-class-switched and class-switched memory B cells in human peripheral blood (Extended Data Fig. 3g). Notably, the relative proportions of B cell subsets in the two donors remained unaffected by data integration, suggesting that integration preserved biological variability (Extended Data Fig. 3h).

To model B cell differentiation as a continuous biological process, we utilized diffusion maps and CellRank^31,32. This revealed that transitional B cells mature into naive B cells and subsequently commit either to the non-class-switched memory B cell fate, such as marginal-zone-like B cells, or differentiate into class-switched memory B cells (Fig. 3f and Extended Data Fig. 3i,j). Next, we queried our B cell maturation atlas to identify driver proteins most strongly associated with differentiation toward the class-switched memory B cell fate (Methods). For instance, transitional B cells migrating from the bone marrow into the circulation displayed rapid downregulation of the pre-B cell receptor, followed by transient decay of CD22, a negative regulator of B cell receptor signaling, and CD185 (CXCR5), which is required for germinal center formation (Fig. 3g and Extended Data Fig. 3g). Lineage commitment was most apparent when visualizing immunoglobulin expression, the hallmark of class-switching, which showed a switch from IgD and IgM toward IgA in class-switched memory B cells, whereas non-class-switched memory B cells retained surface IgD and IgM (Fig. 3h). To identify proteins specifically associated with the class-switched memory B cell lineage compared to the non-class-switched memory B cell fate, we used CytoVI’s generative model to test for differentially expressed proteins. This revealed proteins involved in class-switching, including CD62L, a selectin essential for migration into lymph nodes³³, CD71, a proposed marker for antigen-specific B cells³⁴, and CD100, which has been shown to modulate CD40–CD40L interactions³⁵ (Fig. 3i,j). These results highlight the role of B cell/T cell interactions required for somatic hypermutation and isotype class-switching in B cells. In conclusion, we demonstrate CytoVI’s performance in assembling large-scale protein atlases from conventional cytometry assays that can be used to identify biomarkers or uncover novel regulators of biological processes.

CytoVI enables robust cross-technology integration

To benchmark CytoVI’s performance in integrating datasets generated by distinct antibody-based single-cell technologies, we used CITE-seq data from Stuart et al. and CyTOF data from Levine et al. of human bone marrow samples^36,37. We first harmonized the cell type annotations provided by the original studies and then benchmarked CytoVI’s integration performance against competitor methods using the scIB framework employed before²⁴. CytoVI achieved the overall best performance in this challenging setting, successfully removing technical variation while preserving cell type diversity in its latent space (Extended Data Fig. 4a). FastMNN and Harmony reduced batch effects to some extent but still retained residual technical variation, particularly in the monocyte compartment (Extended Data Fig. 4b). MARIO integrated the matched subset well but, consistent with earlier observations, captured only a fraction of input cells and exhibited pronounced cell type bias, suggesting its use in challenging integration settings where full injective matching is not required (Supplementary Note 7). Cell type ablation experiments revealed that CytoVI was relatively resistant to over-integration compared to competitor methods (Supplementary Note 8). Additionally, CytoVI remained robust under increasing noise added to the protein expression, maintaining competitive cell type preservation and stronger batch correction than FastMNN and Harmony (Extended Data Fig. 4c).

Extended Data Fig. 4. CytoVI performs robust cross-technology integration of different single-cell technologies (cross-technology integration benchmarking).

A: Overview table benchmarking the integration performance of CytoVI, FastMNN, cyCombine, and Harmony for the cross-technology integration of CyTOF (Levine et al., 32 protein markers) and CITE-seq data (Stuart et al., 25 protein markers, 12 protein markers overlap) consisting of human bone marrow cells. Metrics are presented for bio conservation and batch correction, and an aggregated consensus score summarizes the results. B: UMAPs of the unintegrated data space, the CytoVI latent space, the FastMNN embedding, and the Harmony embedding colored by the respective study (left) and cell type (right). C: Heatmaps displaying the aggregated total scores (left panel), bio conservation scores (middle panel), and batch correction scores (right panel) for the indicated methods with increasing amounts of added noise to the data from both technologies.

Together, these cross-technology integration benchmarks show that CytoVI effectively integrates antibody-based single-cell data across different technologies while limiting over-integration.

CytoVI imputes morphological features from mass cytometry data

Unlike flow cytometry, mass cytometry lacks morphological features such as cell size and granularity, which can be important for detecting aberrant cells in hematological settings or providing insights into activation states. To evaluate whether CytoVI could be used to impute forward-scattered light (FSC; a proxy for cell size) and sideward-scattered light (SSC; a proxy for cell granularity) from mass cytometry data, we integrated flow and mass cytometry data from healthy PBMCs^13,14 (Fig. 4a). CytoVI controlled for the technical variation between the two technologies and yielded a technology-harmonized latent space that preserved biological variability (Fig. 4b and Extended Data Fig. 5a). Employing CytoVI’s generative model, we imputed FSC and SSC in the mass cytometry data.

Fig. 4. CytoVI facilitates cross-technology integration and imputation of missing modalities.

a, Schematic overview for the integration of flow (Nuñez et al., 35 protein markers + FSC/SSC) and mass cytometry data (Ingelfinger et al., 36 protein markers, overlap of 17 protein markers). b, UMAP of the CytoVI latent space displaying the integration of mass (yellow) and flow (teal) cytometry data. c, Scatter plot displaying the imputed morphological features FSC and SSC. Cells are colored by the indicated cell types. d, Scatter plot showing the mean FSC and SSC for the indicated immune population measured by flow cytometry (ground truth) and imputed in mass cytometry data. Statistics refer to Pearson’s correlation coefficient and two-sided tests of the null hypothesis of zero correlation. e, Schematic overview of the B cell non-Hodgkin lymphoma analysis. f, UMAP of the CytoVI latent space of lymph node samples of B cell non-Hodgkin lymphoma patients (Roider et al., 12 protein markers + FSC/SSC per panel, 8 protein markers overlap) colored by the antibody panel utilized to generate the data (left) or the indicated lymphoma entities (right). rLN (n = 7); DLBCL (n = 17); MCL (n = 9); MZL (n = 6); and FL (n = 24). g, UMAP of the CytoVI latent space displaying the differential-abundance scores for the indicated disease entities. h, UMAP of the CytoVI latent space highlighting the top-two differentially abundant clusters for each disease entity. i, Violin plots displaying the relative frequencies of the indicated differential-abundance clusters across the different disease entities. P values refer to two-sided pair-wise Mann–Whitney U-tests in relation to rLNs. rLN (n = 7), DLBCL (n = 17), MCL (n = 9), MZL (n = 6) and FL (n = 24). The inner box-and-whisker plots of violin plots display the median as the center, the box bounds the interquartile range (25th–75th percentiles), and the whiskers indicate the range of the data. j, Schematic overview displaying the integration of the flow cytometry cohort with paired CITE-seq data (Roider et al., 70 protein markers, 14 protein markers overlap) for imputation of the transcriptome. k, UMAP of the CytoVI latent space showing the protein expression of the indicated markers measured by flow cytometry (top) and the imputed gene expression of the corresponding transcripts (bottom). l, MA plot ranking the log fold change of imputed gene expression of the double-negative T cell cluster compared to all other T cells in relation to the median normalized gene expression of the double-negative T cell cluster. m, UMAP of the CytoVI latent space showing the imputed gene expression of the indicated transcripts. n, Dot plot displaying the scaled normalized gene expression profile of the γδ T cell cluster detected in the CITE-seq dataset in relation to all other T cell subsets. The color bars in k and m indicate relative expression. T_EM, effector memory T; T_FH, follicular helper T; Mem., memory; Cyt., cytotoxic; DA, differential abundance; Dir, direction. Panels created in BioRender: a, Ingelfinger, F. https://biorender.com/z8uf5xm (2026); e, Ingelfinger, F. https://biorender.com/ntnaivw (2026); j, Ingelfinger, F. https://biorender.com/f3ihitn (2026).

Extended Data Fig. 5. CytoVI imputes morphological features in mass cytometry data.

A: UMAP of the CytoVI latent space displaying the integration of PBMCs measured by mass cytometry (Ingelfinger et al.) and flow cytometry data (Nunez et al.). Cells are colored by the harmonized cell labels between the datasets obtained from the ground truth annotation as defined in Extended Data Figure 1A-B. B: Scatter plots displaying CD14 and the imputed scatter features in the mass cytometry data (left panel). Cells are colored by the harmonized cell labels between the datasets. Violin plots displaying the imputed scatter features for the indicated immune populations in the mass cytometry dataset (right panel). C: Violin plots displaying the measured ground truth scatter features for the indicated immune populations in the flow cytometry dataset (left panel). Scatter plots displaying the measured ground truth scatter features of the flow cytometry data (right panel). Cells are colored by the harmonized cell labels between the datasets.

Imputed light-scatter features in the mass cytometry data recapitulated typical flow cytometry patterns, including high FSC and SSC in myeloid cells and increased size and granularity in memory compared to naive CD4⁺ and CD8⁺ T cells, consistent with previous reports linking these features to mitochondrial biogenesis and remodeling during T cell activation (Fig. 4c,d and Extended Data Fig. 5b,c)³⁸.

In conclusion, we demonstrated that CytoVI can be utilized to integrate distinct cytometry assays and impute features that are available in only one of the technologies.

CytoVI automates the discovery of T cell states associated with nodal B cell lymphomas

CytoVI’s stochastic latent space enables high-resolution differential-abundance analysis across conditions without a priori cell clustering by comparing sample-level posterior distributions (Methods). To investigate whether differential-abundance scores could automatically detect disease-associated cells, we analyzed flow cytometry data from T cell infiltrates in lymph nodes of 63 patients with B cell non-Hodgkin lymphoma, including diffuse large B cell lymphoma (DLBCL), marginal-zone lymphoma (MZL), mantle cell lymphoma (MCL), follicular lymphoma (FL) and resected control lymph nodes (rLNs)³⁹ (Fig. 4e). The dataset comprised two partially overlapping antibody panels (12 proteins each, with an overlap of eight proteins) and was originally analyzed by manual gating. Using CytoVI, we integrated the flow cytometry data from the two panels and imputed non-overlapping proteins (Fig. 4f and Extended Data Fig. 6a,b).

Extended Data Fig. 6. CytoVI automatically detects T cell states associated with disease in B cell non-Hodgkin lymphoma patients.

A: UMAP of the CytoVI latent space of T cells obtained from lymph node samples of B cell non-Hodgkin lymphoma (BNHL) patients measured by flow cytometry (Roider et al.) colored by the batch in which the data was generated. B: UMAP of the CytoVI latent space of lymph node samples of BNHL patients displaying the imputed expression of the indicated protein markers. C: UMAP of the CytoVI latent space of T cells of BNHL patients colored by the cell state annotation analogous to Roider et al. (left panel). Heatmap displaying the column-scaled mean imputed expression profile of the indicated T cell states (right panel). D: UMAP of the CytoVI latent space of T cells of BNHL patients colored by the differential-abundance (DA) clusters obtained from Leiden clustering of the per-entity differential-abundance scores concatenated to the CytoVI latent space. E: Violin plots displaying the relative frequency of the differential-abundance T cell clusters per sample separated by the panel that was used to measure the cells. The dashed line indicates samples that correspond to the same patient but were measured using a different antibody panel. Statistics indicate the intra-class correlation coefficient (ICC) across all clusters and the associated two-sided p value. F: Heatmap displaying the confusion matrix between the cell state annotations analogous to Roider et al. and the differential-abundance clusters. G: Violin plots displaying the relative frequencies of the indicated differential-abundance clusters across the different disease entities. rLN = resected lymph node (n = 7), DLBCL = diffuse large B cell lymphoma (n = 17), MCL = mantle cell lymphoma (n = 9), MZL = marginal-zone lymphoma (n = 6), FL = follicular lymphoma (n = 24). P values refer to two-sided pair-wise Mann–Whitney U-tests in relation to resected control lymph nodes (rLN). H: Heatmap displaying the confusion matrix between the cell state labels of the flow cytometry data of BNHL patients and the nearest-neighbor annotation in the integrated CytoVI latent space for each cell in the CITE-seq data from Roider et al. To estimate the nearest-neighbor labels, a CytoVI model was trained on both technologies. I: UMAP of the CytoVI latent space of flow cytometry data showing the imputed gene expression of the indicated transcripts. J: UMAP of the SCVI latent space of the transcriptome of the CITE-seq dataset highlighting the γδ T cell cluster obtained from Leiden clustering of the SCVI latent space. K: Violin plots displaying the relative frequency of the γδ T cell cluster obtained from Leiden clustering of the SCVI latent space among each patient in relation to the tumor entity (rLN: n = 8; DLBCL: n = 12; MZL: n = 12; FL: n = 11; MCL: n = 8). The inner box-and-whisker of violin plots display the median as the center, the box bounds the interquartile range (25th-75th percentiles), and the whiskers indicate the range of the data.

To enable cross-comparison with the original study, we clustered the CytoVI latent space and used the original authors’ annotation strategy to divide cells into similar T cell subsets³⁹ (Extended Data Fig. 6c). Next, we computed CytoVI’s label-free differential-abundance scores for each T cell to identify cell subsets enriched in the respective disease entities (Fig. 4g). Concatenating these scores with the coordinates of each cell in latent space, followed by graph-based clustering, revealed T cell states with differential-abundance across the interrogated lymphoma entities (Fig. 4h and Extended Data Fig. 6d). Notably, the frequency of these differential-abundance T cell states was highly similar for cells measured across the two distinct antibody panels, validating the applicability of integrated analysis across both panels simultaneously (intra-class correlation coefficient of 0.99, P = 5.97 × 10⁻¹²; Extended Data Fig. 6e).

Next, we matched the differential abundant T cell states with the annotations from Roider et al.³⁹ (Extended Data Fig. 6f). Consistent with the original study, CytoVI identified naive CD4⁺ and CD8⁺ T cell clusters enriched in control lymph nodes, follicular helper T cells enriched in FL, and proliferating T cells enriched in DLBCL and MZL (Fig. 4i and Extended Data Fig. 6f,g). In contrast to the original report of a general proliferating T cell subset, CytoVI resolved this population as predominantly proliferating CD8⁺ T cells, with a smaller fraction of proliferating helper T cells, highlighting the potential of integrated label-free differential-abundance analysis to detect rare disease-associated cell states (Fig. 4g–i and Extended Data Fig. 6B,G). Notably, our label-free differential-abundance analysis detected a small cluster of CD4/CD8 double-negative T cells enriched in patients with DLBCL and MCL that was not reported by Roider et al.³⁹ (Fig. 4i). These double-negative T cells displayed characteristics of activated cytotoxic memory T cells (KLRG1⁺, CD31⁺, CD69⁺, PD-1⁺ and TIM-3⁺), but the protein markers included in the two antibody panels did not allow confident assignment of these cells to a specific T cell subset (Extended Data Fig. 6b).

Thus, we integrated the flow cytometry data of the BNHL cohort with CITE-seq data from a subset of the patients, imputed the transcriptome of T cells in the flow cytometry dataset, and performed differential gene expression analysis of the double-negative T cell cluster to investigate its lineage identity (Fig. 4j). The confusion matrix based on nearest neighbors in the integrated CytoVI latent space showed that cell identities were well preserved between flow cytometry and CITE-seq for most cell types (Extended Data Fig. 6h). CytoVI accurately imputed transcriptomic features from paired CITE-seq data while preserving expected protein–transcript relationships (Fig. 4k and Supplementary Note 9). Differential expression analysis of the imputed transcriptome indicated that the double-negative T cell subset in DLBCL and MCL represents a cytotoxic γδ T cell population, characterized by high expression of effector molecules (GZMA, GZMB, PRF1, GZMK and IFNG) and innate-like receptors (NKG7, KLRC2, KLRC3 and KLRG1), absence of TCR β transcripts, and expression of TCR γ/δ constant chains, with weak enrichment of TRDV1 (Fig. 4l,m and Extended Data Fig. 6b,i,j).

Of note, these Vδ1 memory T cells have been reported to be enriched in the blood and tumors of patients with DLBCL and have been suggested as immunomodulatory targets for therapy⁴⁰. Analysis of measured CITE-seq transcriptomes confirmed a cytotoxic IFNG⁺ γδ T cell cluster with innate-like receptor expression, and showed a similar trend toward enrichment in DLBCL and MCL despite limited statistical power due to the low number of samples analyzed by CITE-seq (Fig. 4n and Extended Data Fig. 6j,k).

Together, these results show that CytoVI can automatically identify disease-associated cell states. By integrating high-throughput cytometry with transcriptome-wide CITE-seq data, CytoVI combined the scale of cytometry with the molecular resolution of single-cell genomics to enable powerful exploratory cell state characterization in preclinical research.

CytoVI facilitates the automated diagnosis of patients with CLL

Flow cytometry is widely used in clinical practice to diagnose hematological malignancies and guide treatment decisions, but technical variation usually necessitates manual gating for each sample separately. We therefore tested whether CytoVI could disentangle biological from technical variation to enable joint analysis and automated diagnosis of patients with chronic lymphocytic leukemia (CLL) via transfer learning (Fig. 5a and Extended Data Tables 1 and 2).

Fig. 5. CytoVI guides the automated diagnosis of hematological malignancies.

a, Schematic overview of the transfer learning task to automatically classify flow cytometry data of patients with CLL using CytoVI (eight protein markers + FSC/SSC, identical panels). b, UMAP of the unintegrated data space of flow cytometry samples of PBMCs of ten patients with CLL and ten non-detectable-disease controls used to train the reference model colored by batch (= individual patient). c, UMAP of the CytoVI latent space of the same reference data controlling for the observed batch covariate. d, UMAP of the CytoVI latent space of ten patients with CLL and ten non-detectable-disease controls (comprising the reference dataset) colored by the corresponding cell type labels obtained from clustering the latent space (left). Heatmap showing the protein expression profile of the indicated cell types (right). e, Scatter plot showing the relative abundance of CLL cells obtained from clustering the CytoVI latent space for patients with CLL high tumor burden (t.b.) (teal; n = 10) and patients without detectable ground-truth disease in the reference dataset (yellow; n = 10) determined by manual gating of an expert hematologist. The P value corresponds to a two-sided Mann–Whitney U-test. f, Scatter plot for the relative abundance of CLL cells per patient (n = 10) obtained from clustering the CytoVI latent space in the reference dataset compared the ground truth annotation obtained by expert gating. Statistics refer to Pearson’s correlation coefficient and two-sided tests of the null hypothesis of zero correlation. g, UMAP of the CytoVI latent space displaying the reference dataset (yellow) and additional unseen query data (gray) that was mapped to the latent space using a forward pass of the trained reference model. h, Scatter plot showing the relative abundance of CLL cells automatically obtained from mapping new unseen query data into the CytoVI latent space and consecutive label imputation. Frequencies are shown for patients with CLL high tumor burden (teal; n = 14; comparable to newly diagnosed patients with CLL), patients with CLL low tumor burden (blue; n = 21; predominantly minimal residual disease patients) and patients without detectable ground-truth disease in the query dataset (yellow; n = 40). Categories on the x axis correspond to the manual annotations from the hematologist. P values correspond to two-sided pair-wise Mann–Whitney U-test in relation to the no detectable disease controls. i, Scatter plot for the relative abundance of CLL cells per patient automatically obtained from mapping new unseen query data into the latent space with consecutive label imputation compared to the ground truth annotation obtained by expert gating. Statistics refer to Pearson’s correlation coefficient and two-sided tests of the null hypothesis of zero correlation. j, Kernel density estimate plot for the log(kappa/lambda) light-chain expression intensity of the automatically detected CLL populations per patient separated by the ground-truth annotation into patients with CLL high tumor burden (teal), patients with CLL low tumor burden (blue) and patients without detectable disease (yellow). k, Heatmap displaying the log(kappa/lambda) light-chain expression intensity, ground truth light-chain restriction, relative frequency of CLL cells and the uncertainty of the label imputation for the automatically detected CLL cells per patient (n = 35). Each column of the heatmap corresponds to a patient. l, Scatter plot displaying the log(kappa/lambda) light-chain expression intensity in relation to the relative frequency of CLL cells for the automatically detected CLL cells per patient (n = 35 patients with CLL; n = 40 non-detectable-disease controls). The dashed line depicts the decision boundary of a decision-tree classifier trained on both features. Statistics refer to mean and s.d. of the AUC-ROC obtained by fivefold cross-validation. Panel a created in BioRender; Ingelfinger, F. https://biorender.com/1qtrfi2 (2026).

Extended Data Table 1.

Clinical characteristics of the CLL cohort

Table indicating the diagnosis for each patient included in the CLL cohort, the presence of the Matutes criteria in the clonal population, the ground truth relative frequency of CLL cells (among B cells and total leukocytes) assessed using expert manual gating, and the light-chain restriction of the CLL clones.

Extended Data Table 2.

Antibodies used for flow cytometry analysis of patients with CLL

Table displaying the antibody panel, including fluorophores, clones, and manufacturers, for the antibodies used to analyze the CLL cohort.

To train a reference model, we selected a dataset comprising ten patients with no detectable disease and ten patients with high tumor burden CLL (>2% of PBMCs). Despite the use of standardized clinical flow cytometry panels, batch effects hampered joint downstream analysis, which was mitigated by CytoVI (Fig. 5b,c and Extended Data Fig. 7a,b). Clustering of the latent space identified CLL cells with a characteristic CD5⁺CD20^dim phenotype; their relative abundance discriminated patients with a high tumor burden with CLL from non-detectable-disease controls and strongly correlated with clinical annotations (Fig. 5d–f and Extended Data Fig. 7c). We next used the trained model and its cell type annotations as a reference to analyze additional diagnostic flow cytometry samples not seen during training. Mapping these query samples into the reference latent space enabled efficient projection while controlling for technical variation (Fig. 5g). Cell types were then assigned by majority vote of nearest-labeled neighbors in latent space (Methods and Extended Data Fig. 7d). Quantification of the imputed CLL cell fraction clearly distinguished high-burden CLL from no-disease controls, correlated well with manual hematologist annotations, and outperformed a uniform manual-gating baseline, although CLL cell frequency alone was insufficient to separate low tumor burden cases from controls without detectable disease (Fig. 5h,i, Extended Data Fig. 7e and Supplementary Note 10). Computing the κ:λ light-chain ratio in automatically classified CLL cells, a hallmark of monoclonality, revealed skewed light-chain usage in CLL cases and strong concordance with hematologist annotations (Fig. 5j,k).

Extended Data Fig. 7. CytoVI can be applied to automate the diagnosis of chronic lymphocytic leukemia patients via transfer learning.

A, B: UMAP of the unintegrated data space (A) and CytoVI latent space (B) of flow cytometry data of ten chronic lymphocytic leukemia (CLL) patients and ten no-detectable-disease controls (comprising the reference dataset) displaying the arcsinh-transformed protein expression of the indicated markers. C: Strip plots showing the relative frequencies of the indicated immune populations of high tumor-burden patients with CLL (teal; n = 10) and no-detectable-disease controls (yellow; n = 10) in the reference dataset. P values refer to the two-sided Mann–Whitney U-test. D: UMAP of the CytoVI latent space of the query dataset obtained via transfer learning displaying the uncertainty of the label transfer. The uncertainty corresponds to the relative frequency of the most abundant label among the nearest neighbors in the reference dataset for each cell in the query dataset based on the CytoVI latent space. E: Strip plots displaying the relative frequencies of the automatically detected immune populations of high tumor-burden patients with CLL (teal; n = 14), low tumor-burden patients with CLL (blue; n = 21), and no-detectable-disease controls (yellow; n = 40) in the query dataset. P values refer to the two-sided Mann–Whitney U-test. F: Scatter-plot showing the log(Igκ/Igλ) of the automatically detected CLL cells in relation to the relative frequency of the CLL cells for each sample in the query dataset. Color corresponds to the diagnostic groups displayed in E. Dot size corresponds to the ground truth relative frequency of CLL cells determined by expert manual gating.

Finally, integrating both the predicted relative frequency of CLL cells and the κ:λ ratio into a decision-tree classifier improved classification performance, enabling robust differentiation of patients with CLL, including those with minimal residual disease (area under the curve (AUC)-receiver operating characteristic (ROC) 0.74 ± 0.09), which remains challenging to distinguish from non-detectable-disease controls even for expert hematologists (Fig. 5k,l and Extended Data Fig. 7f). These findings highlight the potential of CytoVI as a general framework for scalable and standardized analysis of antibody-based single-cell data through transfer learning. By enabling objective and reproducible identification of disease-associated cell populations together with uncertainty estimates, CytoVI may support applications ranging from disease monitoring and treatment-response assessment to broader translational and research settings.

Discussion

The growing number of publicly available cytometry datasets offers a major opportunity to study cellular systems and disease at scale, yet most datasets are still analyzed in isolation due to a lack of computational methods for harmonizing information across studies, technologies and antibody panels. Here, we introduce CytoVI, a probabilistic generative model for the integrative analysis of antibody-based single-cell data. Implemented in scvi-tools¹¹ and interoperable with established single-cell analysis workflows such as Scanpy⁴¹, CytoVI provides a scalable and statistically rigorous framework specifically tailored to the technical properties of cytometry data. We show that CytoVI supports a broad range of integration tasks across cohorts, antibody panels and technologies. By learning an informative latent space, imputing missing protein measurements, enabling automated annotation and supporting differential protein expression and differential-abundance analyses, CytoVI enables the accumulation of biological insights from datasets that would otherwise remain disconnected.

CytoVI also proved effective in clinical settings. Using routine diagnostic flow cytometry, we show that CytoVI can support automated detection of CLL cells and identify clinically relevant features such as immunoglobulin light-chain restriction, despite technical variation between individually measured patient samples. This addresses a central challenge of clinical flow cytometry highlighted by the EuroFlow Consortium⁴². Because CytoVI provides intrinsic uncertainty estimates, it may be particularly valuable in settings where confidence in automated predictions is crucial. Such models could be deployed either through cloud-based solutions or as pretrained local models for efficient inference in privacy-sensitive clinical environments.

CytoVI also expands the utility of existing preclinical cytometry datasets. Because cytometry is destructive and restricted to predefined antibody panels, proteins not measured in the original experiment cannot be assessed retrospectively. We show that CytoVI can generate probabilistic estimates of unmeasured proteins, thereby extending the value of archived datasets. Moreover, integration with paired CITE-seq data demonstrated that CytoVI can combine the scale of cytometry with the molecular resolution of single-cell genomics. In this setting, CytoVI enabled transcriptome-wide characterization of cellular states across millions of cells and helped uncover a rare γδ T cell population in patients with lymphoma that would have remained inaccessible to either technology alone. As new antibody-based single-cell technologies measuring additional modalities emerge, such as spatial protein organization using molecular pixelation⁴³, we anticipate that CytoVI will be capable of integrating these new technologies. CytoVI could empower large-scale plate-based antibody screens, such as those employed by InfinityFlow⁷, to enable community-accessible organ-scale protein atlases that can be queried using transfer learning approaches. The ability to query large-scale cytometry data through probabilistic models represents a transformative step toward standardizing cytometry data interpretation and application. We envision CytoVI as a first step toward a model-centric view for cytometry data, analogous to recent advances in other domains. By continually learning from expanding cytometry datasets, we anticipate that these generative models will improve generalization, robustness and predictive power across studies and clinical applications. These models could provide a foundation for data-driven discovery from antibody-based single-cell measurements, enabling more systematic characterization of cellular states, biological variation and disease-associated phenotypes across diverse experimental and translational settings.

Methods

This study complies with all relevant ethical regulations. The study was approved by the local ethics committee in Basel (BASEC 2024-00006).

The CytoVI model

CytoVI is a probabilistic latent variable model for antibody-based single-cell technologies, such as flow cytometry, mass cytometry or CITE-seq. Data from each of these technologies can be represented as an expression matrix $X\in {\mathbb{R}}^{N\times P}$, in which each element x_np denotes the expression of protein p for cell n among P total proteins and N total cells. In flow and mass cytometry protein expression data are generated via indirect labeling of antibodies that are conjugated to fluorophores or isotopically pure heavy metal ions. Thus, detected protein intensities are relative measurements that heavily depend on the choice of respective reporter molecules coupled to the respective antibody. To account for the relative nature of antibody-based single-cell data and enhance its suitability for visualization and downstream analysis, cytometry data are typically preprocessed using hyperbolic arcsin, logicle or biexponential transformations and feature-wise scaling to yield comparable marker expression ranges^20,44. To provide flexibility for various applications and common downstream tasks, CytoVI accommodates these preprocessing steps, allowing for the incorporation of any of these transformations into the model. Specifically, we assume that x_np is a transformed and/or scaled protein expression matrix. Additionally, each cell can be associated with additional observed sample-level covariates that can be controlled for in CytoVI’s latent representation, such as categorical nuisance factors, representing experimental batches, utilized technologies or antibody panels applied. For simplicity we will describe the case of a one-hot encoded batch identifier s_n ∈ {1, …, S}. We assume that the protein expression x_np for cell n and protein p can be generated from a d-dimensional latent vector z_n, reflecting a cell’s intrinsic state. CytoVI utilizes a variational autoencoder to approximate the posterior distribution p(z_n∣x_n).

The generative model

We begin by describing the generative model of CytoVI. We assume an isotropic Gaussian prior or a mixture of Gaussians (1) with a diagonal covariance matrix for the latent space z, where μ_k, ${\sigma }_k^2$ and π_k are learnable parameters and denote the mean, s.d. and the mixing coefficient for the k-th component, respectively:

$${\mathbf{z}}_n~\left\{\begin{array}{cc}\mathcal{N}\left(0,1\right),\hfill & \text{if isotropic Gaussian prior}\hfill \\ \sum _{k=1}^K{\pi }_k\mathcal{N}\left({\mu }_k,{\sigma }_k^2\right),\hfill & \text{if mixture of Gaussians prior}\hfill \end{array}\right)$$ 1

The choice of the prior depends on the expected biological variation across cells in the corresponding dataset. Homogeneous datasets are best modeled using the isotropic Gaussian prior, whereas the mixture of Gaussian prior is best applied in heterogeneous samples comprising various different cell types. By default, CytoVI models the prior for the latent space z_n as a mixture of Gaussians, which increases the expressiveness of the latent space and has excelled for the integration of large single-cell studies¹². When additional knowledge about the cells present in the dataset is available (for example cells in the dataset have been annotated before), the prior distribution can be conditioned on these labels to weakly inform the latent representations. Specifically, we adjust the mixture component weights by incorporating a one-hot encoded label vector y, such that:

$${\pi }_k^{\prime }={\pi }_k+\lambda \mathbf{y}$$ 2

In this case, λ is a hyperparameter that determines the weight for the existing cell type label on the prior (default λ = 10).

The generative network models the observed protein expression x_np as a function of each cell’s latent vector z_n and its corresponding one-hot batch vector s_n:

$$f_x:{\mathbb{R}}^d\times {\left\{0,1\right\}}^S\to {\mathbb{R}}^P$$ 3

Here, f_x denotes the nonlinear mappings generated by the decoder network. Finally we describe our observation model using either a Gaussian, parameterized by the mean μ_np and s.d. ${\sigma }_{np}^2$, or β distribution, which is parameterized using the shape parameters α_np and β_np (4):

$$x_{np}\mid {\mathbf{z}}_n,{\mathbf{s}}_n~\left\{\begin{array}{cc}\mathcal{N}\left({\mu }_{np},{\sigma }_{np}^2\right),\hfill & \text{if Gaussian model}\hfill \\ \mathcal{B}\left({\alpha }_{np},{\beta }_{np}\right),\hfill & \text{if Beta model}\hfill \end{array}\right)$$ 4

The choice of the protein likelihood distribution is determined by the specific preprocessing steps applied to the data. By default, we employ a Gaussian emission distribution to model cytometry data that has undergone arcsinh or log-transformation. In cases where the data have been scaled to a [0, 1] range with or without prior transformation, a β distribution can be utilized to model the bound values.

Variational approximation and training procedure

As the posterior distribution p(z_n∣x_n) is intractable, we rely on variational inference to learn a posterior approximation q(z_n∣x_n), similar to our previous work⁴⁵. The variational posterior is chosen to be a Gaussian with a diagonal covariance matrix, with parameters given by a multilayer perceptron encoder f_z applied to x_n.

$$f_z:{\mathbb{R}}^P\to {\mathbb{R}}^d$$ 5

During training, the evidence lower bound is optimized with respect to the inference and generative network parameters using mini-batch stochastic gradient descent. The evidence lower bound consists of a reconstruction term, which encourages the reconstruction of the observed data x given z, and the Kullback–Leibler divergence term that regularizes the posterior distribution:

$$\mathrm{ELBO}={\mathbb{E}}_{q\left(\mathbf{z}\mid \mathbf{x}\right)}\left[\log p\left(\mathbf{x}\mid \mathbf{z}\right)\right]-\mathrm{KL}\left[q\left(\mathbf{z}\mid \mathbf{x}\right)\parallel p\left(\mathbf{z}\right)\right]$$ 6

Handling of overlapping antibody panels

Integration of antibody-based single-cell experiments is often constraint by different input features, such as integration of two flow cytometry datasets employing different antibody panels, resulting in missing values in the expression matrix X after concatenation. We therefore apply a masking scheme for the unobserved proteins similarly to what was proposed in our previous work⁴⁶.

Let ${\mathbf{x}}_n^{\left(s\right)}\in {\mathbb{R}}^{P_s}$ represent the observed input features for cell n in batch s, where s = 1, …, S, and P_s denotes the number of features in batch s. The panel set of each batch is denoted ${\mathcal{T}}_s$, such that $\mid {\mathcal{T}}_s\mid =P_s$. We define the set of shared features among all batches as $\mathcal{I}$, where:

$$\mathcal{I}:={\mathcal{T}}_1\cap {\mathcal{T}}_2\cap \cdots \cap {\mathcal{T}}_S$$

In these cases the encoder network processes only the shared features $\mathcal{I}$ to compute z_n:

$${\mathbf{z}}_n~q\left({\mathbf{z}}_n\mid {\mathbf{x}}_n^{\mathcal{(I)}}\right)$$ 7

The decoder network reconstructs the union of the features for each of the batches denoted as $\mathcal{U}:={\mathcal{T}}_1\cup {\mathcal{T}}_2\cup \cdots \cup {\mathcal{T}}_S$. This yields a reconstructed protein expression vector ${\widehat{\mathbf{x}}}_n^{\mathcal{(U)}}$ for each cell. To handle missing values during training, we generate binary feature masks, ${\mathbf{M}}_n^{\left(s\right)}\in {\left\{0,1\right\}}^{P_s}$ for each batch s = 1, …, S, where:

$$M_{np}^{\left(s\right)}=\left\{\begin{array}{cc}1\hfill & \text{if feature}p\in {\mathcal{T}}_s\hfill \\ 0\hfill & \text{otherwise}\hfill \end{array}\right)$$ 8

We utilize ${\mathbf{M}}_{np}^{\left(s\right)}$ to mask the reconstruction loss for features p that were not observed in the respective batch.

$${\mathcal{L}}_{\text{rec}}^{\left(c\right)}=\sum _{p\in \mathcal{(U)}}{\mathbf{M}}_{np}^{\left(s\right)}\times \log p\left(x_{np}^{\left(s\right)}\mid {\widehat{x}}_{np}^{\left(s\right)}\right)$$ 9

To encourage latent space mixing in case of overlapping antibody panels, we include an adversarial classifier that penalizes the discrimination of cells between employed antibody panels, as proposed in TotalVI⁴⁶.

Imputation of missing proteins

When integrating data generated using distinct antibody panels, CytoVI can impute missing proteins. Specifically, if a protein p was not measured in batch s but was observed in batch $s\prime$, we perform counterfactual decoding to infer its expression.

To achieve this, we first sample the latent representation z_n from the approximate posterior:

$${\mathbf{z}}_n~q\left({\mathbf{z}}_n\mid {\mathbf{x}}_{n,\mathcal{I}},{\mathbf{s}}_n=s\right)$$ 10

where x_n,I represents the subset of features shared across batches. Using this latent representation, we generate the missing protein’s expression by decoding while conditioning on batch $s\prime$:

$$x_{np}^{\left(s\right)}~p\left(x_{np}\mid {\mathbf{z}}_n,{\mathbf{s}}_n=s\prime \right)$$ 11

This procedure reconstructs the missing protein expression in batch s using knowledge from batch $s\prime$. A similar counterfactual decoding can also be applied to data that have been generated using the same antibody panel to obtain a stochastic batch-corrected protein expression estimate, analogous to our previous work for single-cell transcriptomics⁴⁵.

Implementation of CytoVI

CytoVI is implemented as a PyTorch-based model within the scvi-tools framework¹¹. While the model exposes hyperparameters for customization, it utilizes a default encoder-decoder architecture, each consisting of a single hidden layer with 128 hidden units. Model parameters are optimized using the Adam optimizer with weight decay. The dimensionality of the latent space z_n is primarily influenced by the biological characteristics of the sample, as well as the number P and quality of protein markers in the dataset. Based on our observations, latent space dimensions d in the range of 10 to 20 typically yield optimal results for batch correction while preserving biological variation. The default heuristic for determining the latent space dimension is $\frac{P}{2}$ and constrained between 10 and 20. Additionally, we set the number of components K in the mixture of Gaussian prior to match the dimension of latent space d by default.

Differential protein expression analysis

CytoVI leverages the trained generative model for testing differentially expressed proteins using the Bayesian framework proposed earlier⁴⁷. For this we consider cells to be stratified into categories c_n that represent our covariate of interest (such as disease state and response to therapy). To identify differentially expressed proteins between cells of the group c_a versus c_b, we first perform uniform sampling with replacement of the posterior distribution of cells among each group. We then estimate the log fold change ${\lambda }_{a,b}^p$ in generated expression ${\widehat{x}}_p$ for protein p between pairs of cells a and b belonging to groups c_a and c_b:

$${\lambda }_{a,b}^p:={\log }_2\left({\widehat{x}}_{a,p}\right)-{\log }_2\left({\widehat{x}}_{b,p}\right)$$ 12

We then estimate the probability that protein p is differentially expressed between cells of c_a and c_b by assessing the fractions of posterior samples for which ${\lambda }_{a,b}^p$ is larger than a predefined effect size threshold δ (defaults to 0.25).

$$p\left(∣{\lambda }_{a,b}^p∣\ge \delta \mathrm{\mid }{c}_a,c_b\right)\approx \int \mathbb{1}\left\{∣{\lambda }_{a,b}^p∣\ge \delta \right\}q\left({\mathbf{z}}_a\mathrm{\mid }{c}_a\right)q\left({\mathbf{z}}_b\mathrm{\mid }{c}_b\right)\mathrm{d}{\mathbf{z}}_a\mathrm{d}{\mathbf{z}}_b,$$ 13

To account for batch effects in differential expression estimation, we perform counterfactual decoding across all possible batch conditions s_n and compute the differential expression probability as the average over these counterfactual samples. As cytometry data are often feature-wise normalized to the [0, 1] range, we apply numerical clipping to expression estimates at the boundaries. Specifically, values of exactly 0 or 1 are adjusted to 0 + ϵ and 1 − ϵ, respectively, where ϵ = 10⁻⁶, to ensure numerical stability in downstream computations.

Label-free differential-abundance analysis

CytoVI incorporates a label-free differential-abundance estimate across conditions of interest that can be utilized to identify cellular states associated with a sample covariate of interest (such as presence of disease) as described elsewhere¹². For this we aggregate the approximate posterior distribution q_v(z∣x) for each sample v across n_v cells:

$$q_v\left(\mathbf{z}\mid \mathbf{x}\right):=\frac{1}{n_v}\sum _{n:v_n=v}q\left(\mathbf{z}\mid {\mathbf{x}}_n\right)$$ 14

Accordingly, we aggregate the posterior densities across all samples of a given covariate in A ⊂ {1, …, V_a} and B ⊂ {1, …, V_b} into q_A(z) and q_B(z) and estimate the ratio of enriched and depleted regions in latent space with respect to the categorical covariate:

$$r_{A,B}\left(\mathbf{z}\right):=\log \frac{\frac{1}{\mid A\mid }\sum _{v\in A}{\mathrm{e}}^{q_A\left(\mathbf{z}\right)}}{\frac{1}{\mid B\mid }\sum _{v\in B}{\mathrm{e}}^{q_B\left(\mathbf{z}\right)}}$$ 15

The resulting differential-abundance scores r_A,B(z) can be utilized for a cluster-free analysis of compositional alterations associated with the covariate of interest.

Label transfer between query and reference datasets

CytoVI facilitates transfer learning by leveraging pretrained reference models, following the base implementation of the scArches class in scvi-tools^11,48. Unseen query data can be projected into the latent space of a trained reference model, enabling batch-corrected embeddings without additional training. To transfer annotations, we perform label propagation via nearest-neighbor search in the shared latent space. Specifically, for each query cell, we identify its k-nearest neighbors (k-NNs; default k = 20) in the reference dataset using the approximate nearest-neighbor search algorithm implemented in PyNNDescent. The query cell is then assigned the most frequent label among its neighbors, with the relative abundance of the majority label serving as an uncertainty estimate.

Imputation of missing modalities between query and reference datasets

CytoVI’s latent representation enables the imputation of additional modalities that were not observed during training. We demonstrate its applicability by imputing transcriptomic profiles in flow cytometry data using paired CITE-seq datasets. Analogous to label imputation, we leverage a shared latent representation of protein expression across both modalities by training a CytoVI model that accounts for technical variability between the technologies. To impute continuous covariates, such as normalized RNA expression from CITE-seq data, we identify for each cell in the flow cytometry data its k-NNs (default k = 20) in the protein latent space of the CITE-seq dataset using PyNNDescent. The imputed gene expression vector for the query cell is then obtained by averaging the expression profiles of its 20 nearest neighbors in the reference dataset.

Model evaluation using flow cytometry, mass cytometry and CITE-seq data

Analysis of fluorescence intensities in homogeneous microspheres

To determine a suitable distribution for the protein likelihood we analyzed the fluorescent intensities of microspheres that either did not bind fluorescent antibodies at all (negative fraction) or bound a homogenous number of fluorescent antibodies. For this, compensation beads (BioLegend) were used to generate single-stain controls using the following fluorescent antibodies: anti-human CD3 Pacific Blue (Clone UCHT1; BioLegend), anti-human TIGIT BUV395 (Clone 741182; BD Biosciences), anti-human CD56 BV421 (Clone 5.1H11; BioLegend) and anti-human CD25 RB780 (Clone 2A3; BD Biosciences). For each stain, one drop of compensation beads was mixed with one drop of negative bead controls, followed by the addition of 200 μl of PBS. From this mastermix, 50 μl were incubated with 0.5 μl of the respective antibody for 15 min at 4 ^∘C. After incubation, 200 μl of PBS was added, and the samples were centrifuged at 500g for 5 min. The supernatant was discarded, and the samples were resuspended in 200 μl of PBS before acquisition on a Cytek Aurora 5L spectral flow cytometer using default settings. Beads were identified based on FSC and SSC using biaxial gating in FlowJo, and positive and negative fractions of antibody-bound beads were exported separately for each staining condition. The resulting .fcs files were imported into Python using the readfcs library. Quantile plots were drawn using Scipy.stats.

Flow cytometry data of healthy donor PBMCs

To evaluate the distribution of fluorescence intensities in flow cytometry data, we analyzed full spectrum cytometry data of cryopreserved PBMCs of a healthy donor that served as the normalization control in a recent SARS-CoV-2 vaccine study¹³. The flow cytometry dataset was downsampled to 100,000 cells and comprised 35 protein parameters and additional morphological features for FSC and SSC. Doublets, dead cells and fluorescent aggregates have been removed using manual gating in FlowJo. The exported .fcs files have been imported into Python using the readfcs library and stored as an AnnData object, similarly as employed for single-cell genomics data. The raw fluorescent intensities of the protein markers (excluding scatter features) were transformed using a hyperbolic arcsin transformation with a global cofactor of 2,000 using CytoVI’s preprocessing module. The hyperbolic arcsin transformed expression values (protein markers and scatter features) were scaled feature-wise to a range of [0, 1]. We next computed the nearest-neighbor graph based on the input feature representation using Scanpy⁴¹. The resulting neighbor graph was used to compute UMAP and subsequently subjected to Leiden clustering using Scanpy. Leiden clusters have been manually annotated into the most abundant leukocyte populations in the peripheral blood based on key protein marker expression profiles. For the benchmarking of batch integration performance, we utilized a technical replicate of the same healthy donor that was cryopreserved and analyzed by a full spectrum analyzer in a different batch, serving as an internal batch control in the original study¹³. The technical replicate has been processed using the same procedure as outlined above to obtain independent annotation of the two replicates into the same leukocyte subsets.

Mass cytometry data of healthy donor PBMCs

To benchmark CytoVI’s ability to model mass cytometry data, we accessed mass cytometry data comprising 36 surface and intracellular protein markers of two healthy donor PBMCs that were used as normalization controls in a recent study of monozygotic twin pairs discordant for multiple sclerosis¹⁴. The two healthy donors were measured in one individual batch using a live cell barcoding approach to mitigate technical variation between the samples. Therefore, the concatenated mass cytometry data of both of the samples was processed in a similar approach as outlined above. In short, the data was downsampled to 100,000 cells, transformed using a hyperbolic arcsin transformation with a global cofactor of 10 and feature-wise scaled to a range of [0,1]. CD138 was excluded from the analysis due to unspecific staining as carried out by the original study. The nearest-neighbor graph has been constructed based on all remaining input features of the mass cytometry data of both concatenated samples and was used for UMAP and consecutive Leiden clustering. PBMCs have been manually annotated into the major leukocyte populations.

CITE-seq data of healthy donor PBMCs

To evaluate CytoVI’s ability to model the protein expression measured using CITE-seq, we accessed CITE-seq data across a panel of 228 surface protein markers of PBMCs from an HIV vaccine trial¹⁵. To evaluate data generated from a single batch and enable a cross-comparison to healthy PBMCs measured by flow and mass cytometry, we selected one individual donor of the CITE-seq dataset at baseline (before any vaccination applied). As the original data contained cell annotations, we utilized these labels as the ground-truth cell-type annotations. Similarly, as for the flow and mass cytometry data, the protein data of the CITE-seq dataset was transformed using a hyperbolic arcsin transformation with a global cofactor of 5 and each feature was scaled to a range of [0, 1].

Posterior predictive checks

To assess the extent to which CytoVI is able to generate protein expression data that closely matches the observed data we performed posterior predictive checks. Posterior predictive checks were performed using healthy PBMCs of flow cytometry data from ref. ¹³, mass cytometry data from ref. ¹⁴ and CITE-seq data from ref. ¹⁵ as outlined above. Histograms of selected markers measured across flow cytometry, mass cytometry and CITE-seq were drawn using CytoVI’s plotting module. For each technology three separate CytoVI models were trained using either (1) a Gaussian protein likelihood operating on the arcsinh-transformed data without additional feature scaling; (2) a Gaussian protein likelihood operating on the arcsinh-transformed and feature-wise scaled data; and (3) a β distribution for the protein likelihood operating on the arcsinh-transformed and feature-wise scaled data. For the mass cytometry data a latent space of size d = 10 was chosen across each of the models, while we chose a latent space of size d = 20 for flow cytometry and CITE-seq data. Posterior predictive checks were performed using an extension of the scvi.criticism module. As a naive baseline, we employed factor analysis using the scikit-learn package, following the implementation described in ref. ⁴⁶. For each model, we generated ten posterior predictive samples from $p\left(\widehat{\mathbf{x}}\mid \mathbf{x}\right)$ and compared the mean coefficient of variation (CV) across both the cell and protein axes to the CV of the observed data. To quantitatively assess model performance, we computed the mean absolute error, as well as Pearson’s and Spearman’s correlation coefficients between each model’s generated CV and the observed CV across both, the cell and protein axes.

Latent-space interpretability

To assess to what extent the CytoVI latent space captured biological variation we utilized the CytoVI models for each of the technologies utilized in the posterior predictive checks. The latent space interpretability was demonstrated for models using a Gaussian protein likelihood that performed best during posterior predictive checks. To analyze the structure of the d-dimensional latent space (d = 10 for mass cytometry and d = 20 for flow cytometry and CITE-seq), we constructed a k-NN graph based on the latent embeddings and applied UMAP using Scanpy. The ground-truth cell-type annotations that were either obtained from the original study if available (for the CITE-seq data) or generated from clustering the data space (see above), were visualized on the UMAP of the latent representation. To assess whether the CytoVI latent space facilitates de novo cell typing, we applied Leiden clustering on the latent representation for the flow cytometry data using Scanpy. The ground-truth cell-type annotations and Leiden clusters obtained from the CytoVI latent space were compared by computing the confusion matrix between the two labels using Scanpy and reordering the matrix using a linear sum assignment implemented in Scipy.

Benchmarking of integration performance

Benchmarking of integration performance was conducted using the annotated flow cytometry data from cryopreserved PBMC replicates¹³. CytoVI was benchmarked against cyCombine, FastMNN and Harmony, integration methods that do not inherently require normalization controls and are therefore not restricted to experiments, where such controls are available^21–23. Integration using cyCombine was performed in R using default parameters. FastMNN, implemented in the batchelor R library, was run with an embedding dimensionality of d = 20 (to account for the lower dimensionality of cytometry data compared to single-cell genomics) and otherwise default parameters. cyCombine’s workflow necessitates an intermediate self-organizing map clustering step, followed by batch correction applied to the original arcsinh-transformed protein expression values. To ensure comparability across methods, we post-processed the cyCombine-corrected expression matrix by applying the same transformation used during label generation (either min-max, z-score or rank scaling) before downstream analysis. Accordingly, CytoVI, FastMNN and Harmony were also evaluated using each of the preprocessing options (min-max, z-score or rank scaling). Despite CytoVI’s assumption of a Gaussian likelihood to model protein expression data (which is incompatible with rank-scaled data), we included the integration of CytoVI on rank-scaled data for completeness of the comparison. As before, we trained the CytoVI models by optimizing latent space of size d = 20 with default parameters. Harmony integration was performed using the Python implementation using default parameters. We evaluated batch correction and biological conservation across integration methods using scvi.scib metrics, employing the metrics defined elsewhere²⁴. Cells that could not be unambiguously assigned during annotation were excluded from the integration benchmarking. Benchmarking was carried out using either the latent embedding or the corrected protein expression matrices from the individual methods across the three preprocessing options with batch and ground-truth cell annotations provided. For computational efficacy, evaluation of integration methods was performed on a random subset of 10,000 cells per batch. Runtime analysis on a random subset of the data (1,000–1,000,000 cells) was evaluated on a computation node equipped with 64 GB of system RAM, a 16-core CPU, and a single NVIDIA A100-SXM4 GPU with 80 GB of memory. To evaluate CytoVI’s sensitivity to feature ablation, we iteratively masked an increasing number of randomly selected markers in one of the batches of the Nuñez et al. dataset, trained a CytoVI model and benchmarked integration performance using scIB.

To compare CytoVI to MARIO, a cell-matching algorithm that consecutively generates a batch-corrected embedding using canonical correlation analysis, we ran the Python version of MARIO using default parameters²⁵. MARIO only matches high confidence cell pairs and may additionally match the same cell from a query batch multiple times with the reference batch. This procedure resulted in integration of only approximately 60% of input cells. To enable comparison with CytoVI, we evaluated CytoVI on the same subset of cells that MARIO could match with high confidence using scIB. For the cross-technology integration benchmark, the fraction of cells that could not be matched by MARIO were more pronounced and did not facilitate benchmarking of label preservation as some cell types remained entirely unmatched. We therefore applied MARIO without its default batching procedure and evaluated each method on the remaining dataset that MARIO could match.

Evaluation of protein imputation performance

To benchmark CytoVI’s performance for the imputation of missing protein markers, we utilized flow cytometry data of one individual PBMC donor from the Nuñez et al. dataset¹³. The flow cytometry data were randomly downsampled to 50,000 cells. CXCR3 and PD-1 were excluded from the analysis due to unspecific staining patterns. The resulting data was split into semi-synthetic batches by randomly assigning cells into one of two batch categories. Next we iteratively masked one marker at the time in one of the batches, trained a CytoVI model to integrate the two semi-synthetic batches and imputed the expression of the masked protein. To obtain robust estimates, we sampled each cell 50 times from the approximate posterior distribution and decoded its protein expression for the masked marker. The mean imputed expression across these 50 posterior samples was then compared to the observed protein expression for the respective cell. This procedure was repeated iteratively until all protein markers had been predicted. To estimate the uncertainty in the imputed protein expression we computed the standard deviation of the decoded protein expression across the 50 samples and assessed the relationship with the prediction error (the absolute difference between generated and observed protein expression) for the 50 samples across both, the cell and the protein axes. The imputation performance of CytoVI has been compared to cyCombine, CyTOFmerge and k-NN imputation, albeit the former one has been proposed to be used solely for visualization purposes due to the density sampling, which may hamper direct comparison^21,49. k-NN imputation has been performed using the scikit-learn implementation with ten nearest neighbors. CyTOFmerge was run with default parameters using a revised implementation that accepts preloaded data instead of raw .fcs files and robustly handles edge cases in partially overlapping panels (see reproducibility repository). For cyCombine, we followed the standard workflow, using default parameters and z-score scaling of the arcsinh-transformed protein expression to generate the self-organizing map labels. Imputation performance of the individual methods for each protein marker was compared by computing the Pearson’s and Spearman’s correlation coefficient and the Wasserstein distance between imputed expression and the held out observed expression. For markers that showed a clear bimodal expression pattern, the imputation performance was evaluated based on the binary marker positivity per cell. For this we fitted a two-component Gaussian mixture model with the protein expression of an individual marker separately for each model and computed binary classification metrics between the observed ground truth marker positivity and marker positivity of each imputation method using scikit-learn. To evaluate the sensitivity of CytoVI’s imputation performance to fluorophore choices, we correlated the Pearson’s correlation coefficient of imputed versus ground-truth marker expression with the relative brightness of the respective fluorophore (extracted from BD’s relative fluorochrome resolution chart) and the maximum channel range.

Imputation of missing markers to increase the dimensionality of cytometry assays

Imputation of missing proteins from two distinct flow cytometry antibody panels

To showcase CytoVI’s performance in integrating flow cytometry data from two distinct studies utilizing different antibody panels, we integrated data of PBMCs obtained from Nuñez et al.¹³ with flow cytometry data of PBMCs of an additional donor from a recent study investigating the systemic immune compartment of patients with SARS-CoV-2 (ref. ²⁶). We used a weakly supervised approach by incorporating the cell type labels to inform the mixture of Gaussian prior of the latent space z. For comparative analysis we also report the integration without cell type-informed priors (Extended Data Fig. 3c). The additional dataset from Kreutmair et al. interrogated 30 surface protein markers and was generated using a full spectrum cytometer²⁶. The flow cytometry data of the Kreutmair et al. study was processed as described above for the other datasets. In short, the dataset was downsampled to 100,000 cells, fluorescent intensities were transformed using a hyperbolic arcsin transformation employing a constant cofactor of 2,000 and feature-wise scaled to a range of [0, 1]. The scaled protein expression was used to construct a nearest-neighbor graph, compute UMAP and perform Leiden clustering using Scanpy. The resulting clusters were annotated into the major leukocyte subsets in the peripheral blood and harmonized to the cell type annotation in the Nuñez et al. dataset¹³. CD1c expression in the Kreutmair et al. data showed high unspecific staining patterns and was excluded from the integration. The concatenated dataset of the two studies was characterized by a shared backbone set of the following overlapping protein markers: CCR4, CCR7, CD123, CD14, CD16, CD19, CD27, CD3, CD38, CD4, CD45RA, CD56, CD8, CD95 and HLA-DR. A CytoVI model was trained using default parameters controlling for the dataset identifier as batch covariate and incorporating the harmonized cell type annotations to weakly inform the prior of the latent space. Similar to before, a neighbor graph was constructed from the joint latent space and used for visualization using UMAP. Missing markers were imputed by sampling ten times from the approximate posterior distribution, generating the decoded protein expression for all markers and computing the mean across the sample dimension. IgG, measured in the Kreutmair et al. displayed unexpected staining patterns. The high IgG expression in NK cells, dendritic cells and monocytes was most likely caused by Fcγ receptor binding of the Human TruStain FcX blocking reagent, which contains human IgG and is not recommended with anti-IgG staining²⁶. To avoid misinterpretation, we removed IgG expression from all affected visualizations.

To assess the sensitivity of CytoVI’s imputation performance to backbone marker ablations, we iteratively masked one of the backbone markers at a time from one of the batches and evaluated the change in imputation performance compared to baseline (no backbone marker ablated). For each ablated backbone marker, we computed the mean squared error drift between the imputation results in the absence of the backbone marker in relation to the baseline (when no backbone marker was ablated) normalized by the baseline variance of that marker.

Generation of an integrated B cell maturation atlas across 350 markers

For the generation of the imputed single-cell B cell maturation atlas, mass cytometry data from Glass et al. were accessed³⁰. This dataset comprises PBMCs of two healthy donors each measured across 12 distinct mass cytometry antibody panels, resulting in 24 unique samples. Single viable B cells were obtained from the mass cytometry .fcs files using manual gating in FlowJo and imported into Python. As in the original publication CD138 was excluded from the analysis due to unspecific staining. The mass cytometry data of the individual samples was transformed using a hyperbolic arcsin transformation with a constant cofactor of 5 and feature-wise scaled to a [0, 1] range. Due to the computationally expensive downstream analysis of trajectory inference methods, the dataset was downsampled to 48,000 B cells in equal proportions from each sample³². A CytoVI model was trained by controlling for the sample key and only encoding the relevant B cell markers that were shared across the antibody panels: CD24, CD27, CD38, IgD and IgM. As the B cell population was expected to comprise a relatively homogenous population, we assumed an isotropic prior for the latent space of size d = 4. The trained CytoVI model was utilized to impute the non-overlapping protein markers by sampling ten times from the approximate posterior distribution, decoding the cells through the generative network to obtain the protein expression samples and computing the mean across the sample axis. Diffusion maps were generated by first computing the top 50 principal components, constructing a nearest neighbors graph based on this representation, and subsequently deriving the diffusion components using Scanpy³¹. Leiden clustering was applied to the nearest-neighbor graph in conjunction with manual annotation to dissect the B cell maturation atlas into the major B cell subsets. Pseudotime of the B cell trajectory was estimated using Palantir⁵⁰, employing five components and otherwise default parameters, and using diffusion pseudotime as implemented in Scanpy (default parameters)⁵¹. Cell fate modeling was performed using CellRank’s pseudotime kernel on a random subset of 20,000 cells of the integrated B cell maturation atlas³². As Palantir and diffusion pseudotime showed highly similar results, the diffusion pseudotime was utilized to construct the pseudotime kernel in conjunction with the Generalized Perron Cluster Cluster Analysis estimator. Terminal and initial states were set using previous knowledge about the underlying B cell biology. Fate probabilities were computed using default parameters and the top 50 protein markers were identified by correlating the individual protein expression with the probabilities of B cells to differentiate toward the non-class-switched memory B cell fate. Visualization of the protein expression dynamics in CellRank has been carried out by fitting a Generalized Additive Model with a Gaussian distribution weighting each cell’s contribution to each trajectory according to its fate probabilities. Differential protein expression between class-switched and non-class-switched memory B cells was performed using CytoVI’s generative model (Methods), while controlling for batch variation.

Cross-technology integration and modality imputation

Benchmarking of cross-technology integration

To benchmark CytoVI’s performance in integrating data generated using distinct antibody-based single-cell technologies, we accessed human bone marrow data measured by CyTOF (Levine et al.) and CITE-seq (Stuart et al.)^36,37. CyTOF data was transformed using a hyperbolic arcsine transformation with a global cofactor of 5. We selected cells from one single donor for which cell type annotations were provided by the original study authors. Protein expression counts of the CITE-seq data were transformed using a hyperbolic arcsine transformation with a global cofactor of 5. The cell type annotations from the original study authors were retrieved from both studies and harmonized to the same level of granularity. The transformed protein expression data from both studies was combined and standard-scaled as our previous benchmarking demonstrated that Harmony and cyCombine require at least z-score scaling for efficient integration. We further benchmarked cyCombine on rank-scaled data, which is specifically recommended for difficult integration tasks with cyCombine. The two datasets shared an overlapping backbone of the following protein markers: CD11c, CD123, CD14, CD16, CD19, CD3, CD34, CD38, CD4, CD45RA, CD8 and HLA-DR. A CytoVI model was trained using a fixed adversarial loss during training and default model hyperparameters. Similarly, as for the integration benchmarking of the flow cytometry replicates from Nuñez et al., we benchmarked CytoVI against Harmony, cyCombine and FastMNN. We chose a dimensionality of d = 10 for FastMNN and otherwise utilized default parameters for each method. As before, we used scIB to benchmark integration performance of the different methods²⁴. To evaluate each method’s sensitivity to noise, we injected increasing amount of noise to the scaled expression of the combined dataset using a Gaussian with mean μ = 0 and s.d. σ ranging from 0.1 to 2.

To evaluate robustness against over-integration across methods, we performed a systematic cell type ablation experiment in the cross-technology integration. In each iteration, one cell type was removed from the CITE-seq dataset before integration, and we assessed whether the method erroneously merged the ablated population across technologies. Integration quality was quantified using two complementary metrics: (1) the integration local inverse Simpson’s index computed on shared (non-ablated) cell types, where higher values indicate better mixing; and (2) the log cross-batch-leakage ratio, defined as the fraction of neighbors (k = 30) from the opposite batch around the ablated population relative to shared cell type controls, where lower values indicate less spurious mixing. We further summarized performance with a composite score (NetScore), defined as $\mathrm{\Delta }{\mathrm{iLISI}}_{\mathrm{shared}}-\mathrm{\Delta }\log {\mathrm{XBN}}_{\mathrm{ablated}}$ relative to the unintegrated baseline, where XBN is the cross-batch-neighbor fraction around the ablated cell type relative to shared controls.

Imputation of morphological features from mass cytometry data

To showcase the capability of CytoVI to integrate distinct antibody-based single-cell technologies with each other, we utilized the annotated and preprocessed mass cytometry data of PBMCs from ref. ¹⁴ and flow cytometry data of PBMCs from ref. ¹³ as described above. The dataset shared a common backbone of the following protein markers: CCR4, CCR7, CD127, CD14, CD16, CD19, CD27, CD3, CD38, CD4, CD45RA, CD56, CD69, CD8, FOXP3, HLA-DR and TCRγδ. The cell type annotations between the two datasets were harmonized by assigning leukocyte subsets to the next coarse cell type annotation level and were used to weakly inform the prior of the combined latent space z. A CytoVI model was trained using default hyperparameters. As described previously, a nearest-neighbor graph was computed based on the latent representation of the trained CytoVI model and utilized for UMAP using Scanpy. We generated ten samples from the approximate posterior distribution, decoded the union of features and computed the mean across the sample axis to obtain the imputed expression of all markers measured across the two studies, including the morphological features FSC and SSC. The mean expression of the imputed FSC and SSC per leukocyte subset in the mass cytometry data was compared to the mean of the measured FSC and SSC in the flow cytometry data by computing the Pearson’s correlation coefficient.

Automated detection of T cell states associated with disease in patients with BNHL

To automatically detect T cell states associated with disease using CytoVI, we accessed raw flow cytometry data from a recent study investigating the T cell landscape in lymph nodes of non-Hodgkin B cell lymphomas³⁹. The dataset comprised 63 lymph node samples that were measured in four batches by two overlapping antibody panels (12 surface and intracellular protein markers per panel) using conventional flow cytometry. Raw flow cytometry data was subjected to manual gating in FlowJo to identify viable singlet T cells and eliminate fluorescent aggregates. The individual batches were imported into Python and individually transformed using a hyperbolic arcsin transformation with a constant cofactor of 500 and scaled to the range [0, 1]. The data of the individual batches were merged into a single AnnData object yielding a shared backbone of the following proteins between the two flow cytometry panels: CD45RA, CD3, FOXP3, CD69, CD8, Ki67, CD4 and PD-1. The resulting dataset was downsampled to 10,000 T cells per patient. A CytoVI model was trained using default hyperparameters controlling for the observed batch covariate. Visualization of the latent space and imputation of non-overlapping markers was performed as described above. To obtain a reference annotation with respect to the original study, Leiden clustering of the CytoVI latent space and manual annotation was performed to identify the same T cell subsets as proposed by the original study authors³⁹. To automatically detect T cell states associated with the different disease entities in a label-free manner, we utilized CytoVI’s differential-abundance module (‘Label-free differential-abundance analysis’ section in Methods). In short, we aggregated the approximate posterior distribution for each patient and computed the log-median ratio of patients within a given disease entity relative to all other patients. Thereby, we retrieved a differential-abundance score for each cell and each disease entity that can be used to detect regions in latent space that are characteristic for the respective disease entity. To identify differentially abundant T cell clusters the differential-abundance scores for each cell and disease entity were concatenated to the latent representation of each cell and subjected to k-means clustering using scikit-learn. The confusion matrix between the identified differential-abundance clusters and the ground-truth cell annotation from the original authors was computed in Scanpy and rearranged based on the linear sum assignment. To evaluate the consistency of differentially abundant T cell clusters between the two applied flow cytometry panels, we computed the intra-class correlation coefficient for the relative frequencies of these clusters, using the pingouin library.

Transcriptome imputation in flow cytometry data of patients with BNHL

To impute the transcriptome of cells measured in the flow cytometry data of the BNHL cohort, we accessed paired CITE-seq data consisting of 70 oligonucleotide-tagged antibodies measured across 16 batches by Roider et al.³⁹. The following overlapping protein markers were present in both technologies: CD244, CD25, CD3, CD31, CD38, CD4, CD45RA, CD69, CD8, CXCR5, ICOS, KLRG1, PD-1 and TIM-3. To obtain batch-corrected protein and RNA expression estimates, the 4,000 most highly variable transcripts were selected and a TotalVI model was trained using default parameters⁴⁶. Denoised and batch-corrected protein and RNA expression was obtained by generating ten samples from the approximate posterior distribution, generating the batch-corrected expression for each sample and computing the mean across the sample axis. Similarly, as for the flow cytometry data, the resulting protein expression was transformed using a hyperbolic arcsin transformation using a constant cofactor of 5 and scaled to a range between [0, 1]. The protein expression of the CITE-seq dataset was merged with the flow cytometry data and standard-scaled. Cell type labels provided with the CITE-seq data were harmonized to the labels generated for the flow cytometry data. A CytoVI model was trained to integrate the flow cytometry and the protein expression data of the CITE-seq dataset using an adversarial loss scaled to 4 and a mixture of Gaussian prior that was weakly informed by the harmonized cell type labels. The integration performance was evaluated by comparing the annotation of each cell in the flow cytometry dataset to the annotation of its nearest-neighbor in the CITE-seq data based on the integrated latent space and computing the confusion matrix between the labels. The joint latent representation was utilized by CytoVI’s imputation module to impute the transcriptome for cells measured by flow cytometry (‘Imputation of missing modalities between query and reference datasets’ section in Methods). Imputation performance was benchmarked by comparing the batch-corrected protein expression in the flow cytometry dataset to its corresponding imputed transcript expression after log₁p transformation and computing Spearman’s correlation coefficient. The protein–transcript relationship in the flow cytometry data was compared to the observed relationship in the CITE-seq data by correlating the denoised and scaled protein expression to the log₁p transformed denoised expression of its corresponding transcript. Imputed transcripts were ranked using the log fold change for cells of the differential-abundance cluster compared to all other T cells in the flow cytometry data using Scanpy and visualized in an MA plot. For the validation of the imputed transcript expression in the differentially abundant T cell cluster, a SCVI model was trained for the measured RNA data from the CITE-seq dataset using default parameters and correcting for the batch covariate⁴⁵. The nearest neighbors graph was constructed based on the SCVI latent space and subjected to Leiden clustering using Scanpy to detect the cluster of γδ T cells and the corresponding gene expression was visualized in relation to the ground-truth labels provided with the CITE-seq data.

Automated classification of clinical flow cytometry data of patients with CLL

Clinical characterization of patients with CLL and ethical approval

The clinical analyses of initial diagnosis and disease course from CLL were obtained from patients treated at the University Hospital of Basel between 2020 and 2023 and the study was approved by the local ethic committee (BASEC 2024-00006). The median age of patients with detectable disease was 64 years and for minimal residual disease (MRD)-negative it was 62 years. Patient characteristics are summarized in Extended Data Table 1.

Flow cytometry data of clinical routine analysis

The clinical analyses of initial diagnosis and disease course in patients with CLL were performed on blood samples using ten-color multiparameter flow cytometry (Lyrics, BD). The samples were analyzed sequentially for each patient as they were presented in the clinic, resulting in technical batches comprising one patient each (Supplementary Note 10). Collected samples were subjected to red blood cell lysis with BD FACS Lysing Solution. Following lysis, cells were washed with phosphate-buffered saline. For immunophenotypic characterization of CLL, a standardized flow cytometry panel was employed, including fluorochrome-conjugated monoclonal antibodies directed against CD45, CD19, CD20, CD5, CD200, CD10, κ and λ immunoglobulin light chains. All antibodies were purchased from BD, with the exception of the kappa and lambda antibodies, which were obtained from Dako (Extended Data Table 2). Flow cytometry data acquisition was performed on a BD FACSLyric (BD Biosciences). Data were collected using FACSDiva software, and fluorescence compensation was adjusted using single-stained controls for each fluorochrome. Fluorescence-minus-one controls were included to establish appropriate gating strategies. At least 50,000 total events were acquired per sample for standard immunophenotyping, whereas for MRD detection, a minimum of 500,000 events was required to achieve a sensitivity of 10⁻⁴. According to the diagnostic procedures, PBMCs from each patient were analyzed by an expert hematologist and classified into three distinct categories based on the abundance of CLL cells: (1) no detectable disease, where no CLL cells were identifiable in the blood; (2) high tumor burden CLL, defined as cases, where CLL cells were present in >2% of total PBMCs, corresponding to the expected clinical phenotype of newly diagnosed patients; and (3) low tumor burden CLL, where CLL cells were present <2% of the total PBMCs, reflecting MRD following therapy. MRD negativity was defined as <1 CLL cell detectable per 10,000 leukocytes (10⁻⁴). Clusters of <20 events were considered undetectable. Within the non-detectable-disease controls, two patients without CLL with only reactive changes in the blood and two patients with non-CLL lymphoma were included. In total, 50 non-detectable-disease controls and 45 positive samples were chosen for downstream analysis using CytoVI (Extended Data Table 1).

Generation of a reference model of patients with CLL

To generate a latent representation of leukocytes in the clinical flow cytometry data while accounting for technical variation between staining protocols and data acquisition sessions, we applied the following preprocessing steps. First, doublets, debris and fluorescent aggregates were removed using FlowJo. The cleaned data were then imported into Python using the readfcs library for further analysis. Similarly, as for the other datasets, the cleaned fluorescence intensity data were transformed using a hyperbolic arcsin transformation and scaled to a range of [0, 1]. To train and annotate a well-defined reference model, ten patients with CLL and ten non-detectable-disease controls were selected. A CytoVI model was trained using this reference dataset by explicitly controlling for the sample key as the batch covariate. A latent space of size d = 5 was optimized, utilizing a k = 5 component mixture of Gaussian prior and encoding all protein markers except the immunoglobulin κ λ light chains into the latent space. The immunoglobulin light chains were not passed to the encoder network to optimize a latent space that does not separate between Igκ- or Igλ-restricted CLL clones. The latent representation of the trained model was used to construct a neighbors graph that has been clustered using Scanpy’s Leiden clustering implementation and manually annotated into the most abundant immune populations, including CD20^dimCD5⁺ CLL cells. Relative frequencies of each immune population among total cells per patient were quantified using a Mann–Whitney U-test. Relative frequencies of CLL cells were compared to the ground-truth frequency determined by hematologists using manual gating by computing Pearson’s correlation coefficient.

Automated patient classification using transfer learning

For the automated annotation of cells and consecutive classification of patient samples, we compiled the remaining dataset into the query dataset that was not used to train the model. For this, we loaded the query data into the trained reference model using the transfer learning approach introduced with scArches into scvi-tools^11,48. Passing the query data through the pretrained inference network yielded the latent representation of each cell without additional training. Cell type labels of cells in the query dataset were imputed using label transfer based on the 20 nearest neighbors in the CytoVI latent space (‘Label transfer between query and reference datasets’ section in Methods). Similarly, as for the reference dataset, the relative cell type abundance in the query data was compared to the hematologists classification using a Mann–Whitney U-test and the frequency of imputed CLL cells per patient was compared to the ground-truth annotation using manual gating by computing Pearson’s correlation coefficient. As a baseline we utilized a uniform manual gate that was applied on the raw normalized expression data to detect the frequency of CLL-like cells. To estimate the immunoglobulin light-chain restriction indicative of a monoclonal population, we computed the log ratio of the scaled Igκ-to-Igλ intensity of the CLL cells of each patient separately. For the differentiation of patients with CLL (low and high tumor burden combined) compared to non-detectable-disease controls, we trained a decision-tree classifier of a depth of three using the relative frequency of the imputed CLL cell labels per patient and the log Igκ:Igλ ratio of the CLL cells as input features using the scikit-learn implementation. The model was evaluated using fivefold cross-validation and the mean and s.d. of the AUC-ROC for the binary classification was reported.

Statistics and reproducibility

The statistical tests utilized are specified in the corresponding figure legends. Sample sizes were determined based on previous experience. Beyond the indicated exclusion criteria for patient samples of the clinical cohort or proteins removed due to technical limitations, no data were excluded from the analyses. Whenever applicable randomization was applied, for example, to split in train/test datasets, query/reference datasets or to downsample cells. The Investigators were not blinded to allocation during experiments and outcome assessment.

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

Online content

Any methods, additional references, Nature Portfolio reporting summaries, source data and Extended data, supplementary information, acknowledgements, peer review information; details of author contributions and competing interests; and statements of data and code availability are available at https://doi.org/10.1038/s41592-026-03224-5.

Author contributions

F.I., I.A. and N.Y. designed the study. F.I. and N.Y. conceived of the model. F.I. implemented the CytoVI software with input from N.L., C.E. and J.M. F.I. performed the analysis. N.L., C.E., A. Bakulin, P.B., D.D., J.D., J.M., T.W., R.Z. and C.C.W. provided intellectual and scientific input. N.L., C.E. and M.K. provided technical support. A. Becker and C.E. implemented analysis methods. C.C.W., D.D. and J.D. provided clinical flow cytometry data and provided clinical input. N.Y. and I.A. supervised the work. F.I., I.A. and N.Y. wrote the paper.

Peer review

Peer review information

Nature Methods thanks Kara Davis, who co-reviewed with Abhishek Koladiya; Matei Ionita; and Fiamma Salerno for their contribution to the peer review of this work. Peer reviewer reports are available. Primary Handling Editors: Rita Strack and Madhura Mukhopadhyay, in collaboration with the Nature Methods team.

Funding

I.A. is an Eden and Steven Romick Professorial Chair, supported by the HHMI International Scholar Award, funded by a European Union ERC advanced grant (no. 101055341-TROJAN-Cell), European Union (IMMEDIATE EU no. 101095540), the MBZUAI/WIS joint program on artificial intelligence, the Deutsche Forschungsgemeinschaft (German Research Foundation) Project ID 259373024 - TRR 167, and the Israel Science Foundation grant no. 1944/22, co-funded by the European Union (ERC, MiTE, 101123436). This research was supported by a Research Professorship grant from the Israel Cancer Research Fund, United States-lsrael Binational Science Foundation, Ministry of Innovation, Science & Technology grant (1001703362), Dwek Institute for Cancer Therapy Research, Moross Integrated Cancer Center, EKARD Institute for Cancer Diagnosis Research, Morris Kahn Institute for Human Immunology, Swiss Society Institute for Cancer Prevention Research, Elsie and Marvin Dekelboum Family Foundation, Lotte and John Hecht Memorial Foundation and the Schwartz Reisman Collaborative Science Program, Canada Brain Research Fund, Dr. Daniel C. Andreae, and the Larry and Judy Tanenbaum Family Foundation. This study was supported by the Center for Immunotherapy at the Weizmann Institute of Science. The work on this project was supported by the Chan Zuckerberg Initiative Essential Open Source Software grants EOSS4-0000000121 and EOSS6-0000000388 (N.Y.). It was also supported by the Chan Zuckerberg Initiative Silicon Valley Community Foundation - Single-Cell Biology Data Insights program (DAF2022-249318) and by National Institutes of Health/National Institute on Drug Abuse (NIDA) award R61DA048444 (N.Y.). F.I. was a recipient of the Koshland prize and was funded by a European Molecular Biology Organization (EMBO) postdoctoral fellowship (ALTF 723-2022) and an MSCA postdoctoral fellowship (101106452; STIC-GBM). F.I. received funding from the Deutsche Forschungsgemeinschaft (German Research Foundation) - SFB-1479 - Project ID 441891347-P23.

Data availability

Raw clinical flow cytometry data from patients with CLL generated in this study are subject to ethical and data-protection restrictions under local governance and are available from the corresponding authors upon reasonable request. Flow cytometry data of healthy PBMCs that served as batch normalization controls for the study by Nuñez et al. have been provided by the authors¹³. Mass cytometry data of healthy PBMCs from Ingelfinger et al. have been provided by the authors¹⁴. CITE-seq data of PBMCs from Hao et al. have been accessed from the National Center for Biotechnology Information’s Gene Expression Omnibus under accession no. GSE164378 (ref. ¹⁵). Flow cytometry data of PBMCs from Kreutmair et al. are at https://doi.org/10.17632/ffkvft27ds.2 (ref. ²⁶). Mass cytometry data of healthy B cells from the peripheral blood generated by Glass et al. are accessible under the Flow Repository accession no. FR-FCM-Z2MA. Raw unprocessed mass cytometry data were provided by the study authors³⁰. Mass cytometry data of human bone marrow samples from Levine et al.³⁶ are on GitHub at https://github.com/lmweber/benchmark-data-Levine-32-dim. Processed CITE-seq data of human bone marrow samples from Stuart et al.³⁷ are available on SeuratData (‘bmcite’). Flow cytometry data of patients with BNHL generated by Roider et al. are at https://doi.org/10.6084/m9.figshare.24915633 (ref. ³⁹). CITE-seq data of patients with BNHL from Roider et al. can be accessed at the European Genome–phenome Archive under accession no. EGAS50000000342. Cell annotations of the CITE-seq data have been provided by the original study authors.

Code availability

Code to reproduce the analysis in this paper is available on GitHub at https://github.com/YosefLab/cytovi-reproducibility. CytoVI is available as an open-source software library integrated at https://scvi-tools.org. The reference implementation of CytoVI is available on GitHub at https://github.com/YosefLab/cytovi-reference-implementation.

Competing interests

F.I., I.A., N.Y. and C.C.W. have a pending patent application related to the clinical applicability of this work. The other authors declare no competing interests.

Extended data

is available for this paper at https://doi.org/10.1038/s41592-026-03224-5.

Supplementary information

The online version contains supplementary material available at https://doi.org/10.1038/s41592-026-03224-5.