In the study of the immune system, researchers use cytometry to measure dozens of proteins on millions of individual cells. To make sense of this data, scientists traditionally perform "gating"—a process of drawing manual boundaries (gates) around clusters of cells on two-dimensional scatter plots to identify specific cell types.
However, this manual approach has become a significant bottleneck. As Hu et al. (2022) note, manual gating is labor-intensive, fails to scale to high-dimensional data, and is inherently subjective. Different experts often produce inconsistent results from the same data, leading to poor reproducibility across laboratories. To solve this, a new wave of research is replacing human "eyeballing" with automated, mathematically rigorous frameworks that identify cell populations with higher precision and speed.
The Subjectivity Problem
The core issue with manual gating is that it forces high-dimensional biological reality into a series of subjective 2D decisions. When a researcher draws a gate, they are making a binary choice about where a cell population ends, often ignoring the "fuzzy" overlapping boundaries common in biology. Ji et al. (2018) and colleagues argue that this introduces significant human bias, as the resulting cell counts depend entirely on the practitioner's interpretation of 2D scatter plots.
Furthermore, manual gating is blind to relationships that exist across three or more dimensions simultaneously. By automating this process, researchers can move beyond simple boxes and capture complex protein expression patterns that do not align with predefined categorical boundaries, as highlighted by Wang et al. (2025).
Probabilistic and Geometric Modeling
One way to automate gating is to treat cell populations as statistical distributions rather than hard-edged shapes. Hejblum et al. (2020) proposed using Bayesian nonparametric frameworks, specifically Dirichlet process mixtures, to automatically determine the number and shape of cell populations. By modeling cells using multivariate skew $t$-distributions, these methods can handle the non-elliptical, "stretched" shapes often found in flow cytometry data without human intervention.
Another powerful approach uses Optimal Transport theory. Instead of looking at cells in isolation, researchers like Inouzhe et al. (2020) treat an entire sample as a high-dimensional probability distribution. By calculating the "distance" between these distributions (using Wasserstein distances), the authors can align different samples and match populations even when technical "batch effects" or biological variability shift the data. This geometric perspective allows for the creation of "templates" that represent healthy or diseased states, which can then be used to classify new patients with high accuracy, as demonstrated in the work on ALS diagnosis by Räuber et al. (2023).
Deep Learning and Computer Vision
Recent breakthroughs have also applied modern computer vision and sequence modeling to the problem. Since manual gating is a visual task, Pattekar et al. (2025) developed the UNITO framework, which transforms protein expression data into 2D density maps and treats gating as an image segmentation problem. By using a U-Net architecture—a standard in medical imaging—the authors achieved human-level accuracy while remaining robust against the noise and debris that typically plague manual analysis.
For even higher dimensionality, researchers are turning to architectures like Transformers and Graph Neural Networks (GNNs):
- Set-Transformers: Lygizou et al. (2024) introduced FCM-Former, which uses attention mechanisms to classify cell lineages in pediatric leukemia, bypassing the need for time-consuming manual preprocessing.
- Explainable GNNs: Zhu et al. (2025) developed Cell Decoder, which embeds known biological relationships (like metabolic pathways) directly into a graph structure. This ensures the model isn't just a "black box" but provides human-readable explanations for why a cell was assigned a specific type.
Beyond Simple Classification
Automated methods are also solving technical artifacts that manual gating struggles to handle. For example, "doublets"—two cells accidentally measured as one—are often discarded as noise. Ionita et al. (2025) showed that automated multivariate tools can not only detect these doublets more accurately than humans but can also "deconvolute" them to study how different cells interact.
By shifting from manual art to automated science, these tools enable high-resolution analysis that uncovers rare cell populations previously missed by human eyes. As Mosmann et al. (2024) demonstrated in vaccine trials, automated clustering can reveal complex immune responses that correlate with clinical outcomes, providing a level of detail that manual gating simply cannot reach.
Key Foundational Work
A Survey of Flow Cytometry Data Analysis Methods (2009) — contributed 3 concept elements to this direction.
Merging Mixture Components for Cell Population Identification in Flow Cytometry (2009) — contributed 1 concept elements to this direction.
Swift: Scalable weighted iterative sampling for flow cytometry clustering (2010) — contributed 1 concept elements to this direction.