AI-Assisted Gating: How LLMs Can Accelerate Flow Cytometry Analysis

Gating a flow cytometry dataset is part science, part art, and part institutional memory. An experienced immunologist can look at a scatter plot and immediately identify lymphocyte populations, set appropriate boundaries, and name the resulting gates. But that expertise takes years to develop, and even experts can disagree on gate placement.

What if an AI could assist — not replace — that expertise?

In Cytogence FCS, we’ve built an AI copilot that does exactly that. It doesn’t gate for you. It gates with you.

What the AI Copilot Actually Does

The copilot is not a black box that takes your FCS file and returns a gating hierarchy. That approach fails because gating is context-dependent — the “right” gates depend on your experimental question, your panel design, your tissue type, and your scientific judgment.

Instead, the copilot operates as an interactive assistant with a growing set of specialized tools:

Analysis Tools

list_parameters — List all channels and markers in the loaded file
get_sample_metadata — Retrieve instrument type, acquisition date, event count
get_statistics — Compute population statistics (mean, median, geometric mean MFI, CV) for any parameter within any gate
analyze_distributions — Characterize the distribution shape of any channel

Gating Tools

create_gate — Draw rectangle, polygon, ellipse, or range gates with specified coordinates
evaluate_gate — Check population counts and percentages for any gate
list_gates — Display the current gate hierarchy
delete_gate, rename_gate — Modify existing gates
run_auto_gate — Trigger unsupervised clustering (K-means, DBSCAN, or GMM)

Intelligence Tools

suggest_gating_strategy — Propose a multi-step gating workflow based on the panel and experiment type
name_populations — Suggest biologically meaningful names for gated populations
set_plot_parameters — Change axis assignments for visualization
set_transform — Apply data transforms (linear, log, logicle, arcsinh, hyperlog)

Statistical Tools

run_statistical_test — Execute t-tests, ANOVA, Kruskal-Wallis, or Mann-Whitney U tests between populations, depending on data distribution and experimental design
get_compensation_matrix — Retrieve the current spillover matrix

Three Modes of Assistance

The copilot operates in three modes, each designed for a different use case:

Copilot Mode

The default. You’re analyzing data, and the AI observes your workflow in real time. It might suggest: “You’ve gated on CD3+ cells. Consider creating a CD4/CD8 subset gate next — your panel includes both markers.” You approve or dismiss each suggestion.

Tutor Mode

For trainees and students. The AI explains why each gating step is important, what the populations mean biologically, and what to watch for. It’s like having an experienced flow cytometrist looking over your shoulder.

Designer Mode

For panel design and experiment planning. Before you’ve collected any data, the AI can help you think through your gating strategy based on your marker panel, suggest potential issues (spectral overlap, missing dump channels), and recommend controls.

How It Works Under the Hood

Multi-Provider Architecture

The copilot supports five LLM backends:

Anthropic Claude — Our recommended option for reasoning quality
OpenAI GPT-4 — Widely available alternative
AWS Bedrock — Enterprise cloud deployment with SigV4 authentication
Ollama (local) — For air-gapped environments where data cannot leave the machine
Cytogence Managed — Our hosted endpoint for users who don’t want to manage API keys

Users choose their provider based on their privacy requirements and institutional policies. For labs working with sensitive data, the local Ollama option ensures that no data — not even parameter names — leaves the local machine.

Tool Approval

Every action the AI proposes goes through an approval step. When the copilot suggests creating a gate, you see the proposed coordinates and can approve, modify, or reject. The AI never modifies your data without your explicit consent.

This is a deliberate design choice. Flow cytometry gating has real consequences for downstream analysis, and we believe the human must remain in control of every decision.

Streaming Responses

AI responses stream in real-time — you see the copilot’s response as it develops, not after a long wait. This makes the interaction feel conversational rather than transactional.

What Works Well

Strategy Suggestions

Given a panel and an experiment type (immunophenotyping, intracellular cytokine staining, cell cycle analysis), the copilot can outline a logical gating sequence. For a standard T cell panel, it might suggest:

Time gate (remove acquisition instabilities)
Singlet gate (FSC-A vs. FSC-H)
Live/dead discrimination
Lymphocyte gate (FSC-A vs. SSC-A)
CD3+ gate
CD4/CD8 subsets
Functional markers within each subset

This isn’t novel for an experienced user — but it’s invaluable for trainees, for unfamiliar panels, or for documenting the rationale behind your gating strategy.

Population Naming

After unsupervised clustering identifies distinct populations, the copilot can suggest biologically meaningful names based on marker expression patterns. A cluster that’s CD3+CD4+CD25+FOXP3+ may be suggested as “Regulatory T cells” — saving the manual step of interpreting each cluster’s phenotype.

Statistical Guidance

The copilot can recommend appropriate statistical tests based on your experimental design: paired vs. unpaired, parametric vs. non-parametric, and whether multiple testing correction is needed.

What Doesn’t Work (Yet)

Complex Manual Gates

The AI can suggest gate coordinates, but it can’t yet match the nuance of an expert manually adjusting a polygon gate based on subtle population features visible in a density plot. Automated gate placement is a starting point, not the final answer.

Novel Panels

For well-characterized panels (standard immunophenotyping, common cytokine panels), the copilot’s suggestions are strong. For novel or highly custom panels, it may default to generic suggestions that miss the specific biology you’re studying.

Quantitative Precision

Like all LLMs, the copilot can make errors when reporting exact numbers. Always verify statistics against the actual computed values displayed in the application — don’t rely on the AI’s paraphrasing of your results.

The Privacy Question

Flow cytometry data can be sensitive — patient samples, unpublished research, proprietary panel designs. We designed the AI integration with this in mind:

Local models (Ollama): Zero data leaves your machine. Full local functionality with complete privacy.
Session-only API keys: Keys are held in memory only and zeroed on application close — never written to disk.
Privacy modes: Standard, Enhanced, and Strict modes control what context is sent to cloud providers.
No training on your data: We do not use your queries or data to train any models.

Why This Approach Works

The key design decision is separation of concerns: the AI handles interpretation and suggestion, while the analysis engine executes computations deterministically.

This avoids a common failure mode in AI-driven tools, where models attempt to perform calculations directly and introduce errors. In Cytogence FCS, the AI suggests actions — but all statistical calculations and gating operations are executed by the Rust analysis engine, ensuring correctness even if the AI’s suggestion needs adjustment.

Looking Ahead

AI-assisted flow cytometry is in its early stages. The current copilot accelerates routine workflows and reduces the barrier to entry for new users. As models improve and we refine the tool integration, we expect the copilot to become increasingly capable at handling complex, multi-panel experiments and providing publication-ready analysis summaries.

But the principle won’t change: the AI assists, the researcher decides.

Cytogence FCS is available for Windows, macOS, and Linux. Request a demo to see the AI copilot in action.