CytoVI: Deep generative modeling of antibody-based single cell technologies

Florian Ingelfinger, Nathan Levy, Can Ergen, Artemy Bakulin, Alexander Becker, Pierre Boyeau, Martin Kim, Diana Ditz, Jan Dirks, Jonas Maaskola, Tobias Wertheimer, Robert Zeiser, Corinne C. Widmer, Ido Amit, Nir YosefbioRxiv 2025 · DOI

TL;DR

CytoVI is a new deep learning framework that integrates and analyzes data from various antibody-based single-cell technologies (like flow cytometry and mass cytometry). It tackles challenges like technical noise and missing data, enabling more robust and comprehensive biological insights and even automating clinical diagnoses.

What Problem Does This Paper Address?

Single-cell technologies using antibodies, such as flow cytometry and mass cytometry, are incredibly powerful for understanding cell states, diagnosing diseases, and advancing research. However, analyzing data from these technologies is tricky. Key challenges include:

  • Technical Variation: Differences in experimental protocols, batches, and even the specific technology used can introduce noise and bias, making it hard to compare data.
  • Limited Antibody Panels: Each experiment uses a specific set of antibodies, meaning not all proteins are measured in every cell or study. This limits our ability to combine knowledge across different experiments or panels.
  • Data Integration: The combination of these issues makes it difficult to build large, comprehensive atlases of cell types or to track changes over time or across different conditions.

The authors aim to solve these problems by developing a unified computational framework that can handle these variations and limitations, enabling more statistically rigorous and integrative analyses.

Methods

CytoVI is a probabilistic generative model built on a variational autoencoder (VAE) architecture. Here's a breakdown of its key components:

  • Encoder: This part takes the protein expression data (from flow cytometry, mass cytometry, or CITE-seq) and maps it into a low-dimensional latent space. This latent space is designed to capture the cell's intrinsic biological state, while being corrected for technical variations (like batch effects).
  • Decoder: This part reconstructs the protein expression from the latent representation. It can also be used to impute missing protein measurements – essentially predicting what the expression of a protein would have been if it had been measured.
  • Probabilistic Nature: CytoVI models the uncertainty in its predictions, which is crucial for reliable downstream analysis and decision-making.
  • Handling Overlapping Panels: CytoVI uses a masking scheme to effectively integrate data from experiments that used different antibody panels, imputing the missing measurements.
  • Transfer Learning: CytoVI can be trained on a reference dataset and then used to analyze new, unseen data without retraining, which is particularly useful for clinical applications.

Datasets

The authors evaluated CytoVI on several publicly available datasets, including:

  • PBMC datasets: From flow cytometry, mass cytometry, and CITE-seq technologies, used for model evaluation and demonstrating integration capabilities.
  • B cell maturation atlas: Mass cytometry data from healthy donors, used to build a large-scale protein atlas.
  • B cell non-Hodgkin lymphoma cohort: Flow cytometry data from patients, used to identify disease-associated T cell states.
  • Chronic Lymphocytic Leukemia (CLL) patients: Clinical flow cytometry data, used to demonstrate automated diagnosis via transfer learning.

Key Insights and Ideas

  • Unified Framework: The ability to integrate diverse antibody-based single-cell technologies (flow, mass, CITE-seq) into a single analytical framework is a major step forward.
  • Imputation Power: CytoVI's ability to accurately impute missing protein measurements is critical for overcoming the limitations of restricted antibody panels and building richer datasets.
  • Clinical Translation: The successful application of CytoVI for automated diagnosis of CLL patients highlights its potential to bridge the gap between research and clinical practice.
  • Uncertainty Estimation: By providing estimates of imputation uncertainty, CytoVI offers a more nuanced and trustworthy analysis, allowing researchers to assess the reliability of predictions.

Key Results

  • Outperforms Existing Methods: CytoVI demonstrated superior performance compared to existing computational methods for integration, denoising, and imputation tasks.
  • Integrated B Cell Atlas: CytoVI successfully constructed a comprehensive B cell maturation atlas spanning 350 protein markers.
  • Disease-Associated T Cell States: In BNHL patients, CytoVI identified T cell states linked to disease, which were not easily discernible with traditional methods.
  • Automated CLL Diagnosis: CytoVI achieved high accuracy in classifying CLL patients, even distinguishing between low and high tumor burden, and showed strong concordance with expert manual gating.
  • Morphological Feature Imputation: CytoVI could impute morphological features (like cell size and granularity) from mass cytometry data, which are typically absent.

Limitations and Future Work

  • Computational Cost: While optimized, training deep generative models on large datasets can still be computationally intensive.
  • Data Quality: The performance of CytoVI, like any model, is dependent on the quality of the input data. Technical variations that are not well-captured by the model could still impact results.
  • Future Directions: The authors envision CytoVI as a foundation for a "model-centric view" of cytometry data, where continuous learning from new datasets will improve generalization and robustness. They suggest integrating novel modalities like molecular pixelation and expanding its use for biomarker discovery and personalized medicine.

Summary updated

Abstract

Abstract Due to their robustness, dynamic range and scalability, antibody-based single cell technologies, such as flow cytometry, mass cytometry and CITE-seq, have become an irreplaceable part of routine clinics and a powerful tool for basic research. However, their analysis is complicated by measurement noise and bias, differences between batches, technology platforms, and restricted antibody panels. This results in a limited capacity to accumulate knowledge across technologies, studies, experimental batches, or across different antibody panels. Here, we present CytoVI - a probabilistic generative model designed to address these challenges and enable statistically rigorous and integrative analysis for antibody-based single cell technologies. We show that CytoVI outperforms existing computational methods and effectively handles a variety of integration scenarios. CytoVI enables key functionalities such as generating informative cell embeddings, imputing missing measurements, differential protein expression testing, and automated annotation of cells. We applied CytoVI to generate an integrated B cell maturation atlas across 350 proteins from a set of smaller antibody panels measured by conventional mass cytometry, and identified proteins associated with immunoglobulin class-switching in healthy humans. Using a cohort of B cell non-Hodgkin lymphoma patients measured by flow cytometry, CytoVI uncovered T cell states that are associated with disease. Finally, we show that CytoVI is a robust probabilistic framework for the analysis of standard diagnostic flow cytometry antibody panels, enabling the automated detection of tumor populations and diagnoses of incoming patient samples. CytoVI facilitates accurate and automated analysis in both preclinical and clinical settings and is available as open-source software at scvi-tools.org .
BibTeX
@misc{ingelfinger2025cytovi,
  author = {Florian Ingelfinger and Nathan Levy and Can Ergen and Artemy Bakulin and Alexander Becker and Pierre Boyeau and Martin Kim and Diana Ditz and Jan Dirks and Jonas Maaskola and Tobias Wertheimer and Robert Zeiser and Corinne C. Widmer and Ido Amit and Nir Yosef},
  title = {{CytoVI: Deep generative modeling of antibody-based single cell technologies}},
  year = {2025},
  doi = {10.1101/2025.09.07.674699},
  howpublished = {bioRxiv},
}
Figure