Black Box or Better Insight? Making Sense of AI in Cytometry

Artificial intelligence has rapidly become one of the most discussed technologies in life science research. In flow cytometry, AI and machine learning promise to help researchers analyze increasingly complex, high-dimensional datasets, uncover rare populations, and reduce the time spent on manual gating. Despite the excitement, many cytometrists remain skeptical—and for good reason.

When an algorithm identifies a cell population, how can you be sure it's biologically meaningful? If a machine learning model reaches a conclusion, can you understand how it got there? And perhaps most importantly, can you trust the results enough to publish, validate, or build future experiments around them?

These concerns often stem from a common issue in modern AI: the "black box" problem.

The future of cytometry analysis will not be determined by how much AI is used, but by how transparently it is applied. The goal should not be replacing scientific judgment with algorithms. Instead, AI should help researchers gain deeper insights while preserving the rigor, reproducibility, and interpretability that good science requires.

The Growing Role of AI in Flow Cytometry

Modern flow cytometry experiments routinely generate data with 20, 30, or even 40+ parameters per cell. As panel complexity increases, traditional manual gating becomes increasingly challenging.

Manual analysis remains valuable, particularly when researchers have well-defined hypotheses and established marker strategies. However, manual workflows can become time-consuming, subjective, and risk missing important findings when datasets become highly dimensional. To address these challenges, researchers have increasingly adopted machine learning and AI-driven analysis tools.

Unsupervised clustering such as FlowSOM, Phenograph, SPADE, and X-shift automatically group cells into populations based on similarities across multiple markers. These approaches can reveal cell subsets that may be difficult or impossible to identify through conventional gating strategies. They are particularly useful for discovery-based research and high-dimensional immunophenotyping studies. 

Dimensionality reduction methods such as t-SNE and UMAP help visualize complex datasets by projecting many dimensions into two-dimensional plots. While these techniques are not classification algorithms themselves, they are commonly used alongside clustering methods to explore population structure and identify patterns within datasets.

Neural networks and deep learning architectures are becoming more common in both flow and imaging cytometry applications. These models can identify highly complex patterns that may not be apparent through traditional statistical methods. Deep learning has shown particular promise in image-based cell classification, morphology analysis, and multimodal datasets.

Collectively, these tools have expanded the analytical capabilities of cytometry researchers. However, not all AI approaches offer the same level of transparency.

Understanding the "Black Box" Problem

The term "black box" refers to algorithms that generate outputs without providing clear explanations of how those outputs were produced. In practice, this means researchers may receive a result—such as a cluster assignment, classification, or prediction—without understanding which features drove the decision or whether the result aligns with known biology.

Deep Learning Models

Deep neural networks are often considered the quintessential black-box algorithms.

These models may contain thousands or millions of adjustable parameters distributed across multiple computational layers. While they can achieve impressive predictive performance, it is often difficult to determine exactly why a particular decision was made.

For cytometry researchers, this lack of interpretability can create several challenges:

• Difficulty validating biological relevance
• Reduced confidence in unexpected findings
• Challenges explaining results to collaborators and reviewers
• Concerns about reproducibility across datasets
• Potential barriers in regulated or clinical environments

Complex Ensemble Models

Some advanced machine learning approaches, including certain ensemble methods and proprietary AI platforms, can also become difficult to interpret. Although these models may outperform simpler methods statistically, understanding the reasoning behind their predictions can be challenging.

Visualization Doesn't Equal Explanation

Even commonly used techniques such as t-SNE and UMAP can create misunderstandings.

While these visualizations help researchers explore data, they do not necessarily explain why populations appear where they do on a plot. Two clusters that appear close together visually may not actually be biologically similar, and visual patterns alone should not be treated as mechanistic explanations.

Why Transparency Matters in Cytometry

Unlike consumer applications where prediction accuracy may be the primary objective, scientific research demands explainability.

Researchers must be able to answer fundamental questions:

• Which markers contributed to this result?
• Can the findings be reproduced?
• Does the output align with known biology?
• How sensitive are the results to parameter changes?
• Can another laboratory independently validate the analysis?

Scientific discovery depends on traceability. If an analytical method cannot be understood or explained, researchers may hesitate to trust its conclusions regardless of how sophisticated the underlying algorithm may be.

This is particularly important in translational research, clinical studies, and regulated environments where analytical decisions must often be documented and justified.

The challenge, therefore, is not whether AI should be used in cytometry. The challenge is ensuring that AI enhances scientific understanding rather than obscuring it.

terraFlow's Approach: Transparent AI for Scientific Discovery

At terraFlow, we believe AI should function as a scientific assistant—not an inscrutable decision-maker. Instead of asking researchers to simply trust algorithm outputs, terraFlow is designed to make analytical workflows visible, reproducible, and interpretable.

Human-in-the-Loop Analysis

terraFlow keeps researchers involved throughout the analytical process. Our auto-gating feature gives the user full control to adjust gates after the AI sets thresholds. Gates can be adjusted concurrently for the entire set of samples within an experiment, per group or batch of samples, and for each individual sample. Whichever method the researcher decides to use, they get the final decision about using AI-determined thresholds or setting them manually.

Traceable Analytical Workflows

Every analytical step should be understandable and reproducible. Rather than producing results through hidden processes, terraFlow emphasizes workflows that allow users to see how populations were identified, how analyses were performed, and how conclusions were reached. 

To accomplish this, we run an exhaustive combinatorial analysis on all markers present within a dataset. This process is similar to what a human with limitless time and energy might do when analyzing data with manual gates. There is no black box algorithm, only straightforward marker combinations and a search for cells within these gates. For every phenotype terraFlow discovers within the data, there is a gating pathway that the researcher can use to verify the result with manual gating. This traceability is critical for publication, collaboration, and regulatory confidence.

Biological Context Matters

One limitation of many AI systems is that they optimize mathematical objectives without considering biological meaning. terraFlow is built around the principle that biological interpretation remains central to data analysis. We use an MLL model to provide researchers with biological context in the form of a literature search for all identified phenotypes. The result is a list of relevant publications with highlighted excerpts that mention the identified phenotype and help researchers answer the important questions:

• What does this mean? 
• What am I missing? 
• What do I do next?

The Future of AI in Cytometry

As cytometry datasets continue to grow in complexity, AI will become an increasingly important component of data analysis. The question is no longer whether machine learning belongs in cytometry—it clearly does.

The more important question is what kind of AI researchers should trust.

The future belongs to analytical platforms that combine computational sophistication with scientific transparency. Researchers need tools that accelerate discovery while preserving the ability to understand, validate, and reproduce results. 

When implemented transparently, AI is not a black box. It becomes a powerful lens through which researchers can better understand complex biological systems, uncover meaningful insights, and advance scientific discovery with greater confidence.

References and Further Reading

1. Van Gassen S, Callebaut B, Van Helden MJ, et al. FlowSOM: Using Self-Organizing Maps for Visualization and Interpretation of Cytometry Data. Cytometry Part A. 2015;87(7):636-645. 
2. Becht E, McInnes L, Healy J, et al. Dimensionality Reduction for Visualizing Single-Cell Data Using UMAP. Nature Biotechnology. 2019;37:38-44. 
3. Xu Y. Machine Learning for Flow Cytometry Data Analysis. 2023. 
4. Bini L, Nassajian Mojarrad F, Liarou M, et al. FlowCyt: A Comparative Study of Deep Learning Approaches for Multi-Class Classification in Flow Cytometry Benchmarking. 2024.
5. Kang AS, Kang LC, Mastorides SM, et al. Machine Learning Approaches to Automated Flow Cytometry Diagnosis of Chronic Lymphocytic Leukemia. 2021. 
6. Barsky LW, Black M, Cochran M, et al. International Society for Advancement of Cytometry (ISAC) Flow Cytometry Shared Resource Laboratory (SRL) Best Practices. Cytometry Part A. 2016;89(11):1017–1030. DOI: 10.1002/cyto.a.2301 
7. International Society for Advancement of Cytometry (ISAC). Data Standards: FCS, Gating-ML, and MIFlowCyt Standards.