Unified Probabilistic Analysis of CyTOF: A Deep Generative Approach using CytoOne

Abstract

Extracting meaningful biological signals from Cytometry by time-of-flight (CyTOF) data remains challenging due to heterogeneity, data characteristics, and the presence of various technical artifacts. Current analysis workflows typically rely on task-specific tools assembled into pipelines, which often make inconsistent distributional assumptions and fail to fully leverage the structure of the data. We present CytoOne, a unified probabilistic framework tailored for CyTOF data that integrates batch correction, differential analysis, and visualization within a single model. CytoOne is built upon a Bayesian hierarchical architecture inspired by Nouveau Variational Autoencoders (NVAE) and employs a novel quasi zero-inflated softplus-normal (QZIPN) likelihood to flexibly model the sparse and noisy nature of CyTOF measurements. We demonstrate via qualitative and quantitative evaluations that CytoOne effectively approximates both marginal and joint distributions of CyTOF data, removes batch-specific artifacts, enables fine-grained differential expression analysis, and facilitates interpretable embeddings for exploratory analysis.