The New Wave of AI for Flow Cytometry: A Practitioner's Field Guide

Two years ago, “AI for flow cytometry” mostly meant a clustering algorithm bolted onto the end of an otherwise manual pipeline. Today it spans automated gating tools, pretrained representation-learning models, and the first agentic copilots that take instructions in plain English. Some of it genuinely changes how we work. Some of it is a benchmark number that won’t survive contact with your messy, real-world panel. Telling the two apart has quietly become part of the job.

What follows isn’t a catalog of every paper or model. It’s a practitioner’s view of which ideas seem genuinely useful, which ones still need validation, and where the field appears to be heading.

Why anyone is automating gating in the first place

Manual gating is the bottleneck everyone already knows about. It’s slow, it leans heavily on institutional memory, and — the uncomfortable part — two experienced people gating the same file can draw different gates and report somewhat different population frequencies, particularly for rare or ambiguous populations. That subjectivity is tolerable for a handful of samples on a familiar panel. It stops being tolerable when you’re processing hundreds of files, or when a 30-plus-marker spectral panel makes the manual gating hierarchy genuinely hard to reason about. And that’s before you’ve handled the unglamorous prep — compensation, spectral unmixing, and QC — where conventional algorithms, instrument software, and human oversight still do most of the heavy lifting.

So the pressure to automate is real. But the useful question is never “should we use AI here” in the abstract. It’s “which tool, for which step, validated how.”

Wave 1: automated gating and annotation (the established stuff)

The mature end of the field is unsupervised clustering and automated annotation. FlowSOM (self-organizing maps) and FAUST are workhorses for discovering and labeling populations; biology-driven models like Scyan and training-data approaches like CyAnno fold in prior knowledge. A complementary step is naming: CytoPheno (2025) takes populations you’ve already clustered and assigns descriptive, human-readable cell-type names from their marker patterns — a labeling aid, not a replacement for clustering or gating. On the gating side, methods that learn from expert-curated gates can place new ones automatically.

These are useful and, importantly, generally easier to interrogate than deep neural approaches — you can usually see what the algorithm did and why. Their limits are also well understood: clustering results shift with preprocessing and parameter choices, and a method tuned on one tissue or instrument rarely transfers cleanly to another. We treat this whole category as a fast first pass that an analyst still checks, not as an answer.

Wave 2: the “foundation model” push

The newer, more ambitious work borrows the playbook that reshaped text and images: pretrain a model on a lot of data, then adapt it to specific tasks. Two related threads matter here. One is self-supervised pretraining to get better, more transferable immunophenotyping. The other tackles a problem unique to cytometry — every lab’s panel is different (different markers, orderings, instruments), so a model that assumes a fixed input can’t cope, which is why some of the most interesting work is explicitly panel-agnostic.

Three recent papers stake out related parts of this space from different directions:

• cyMAE (Cytometry Masked Autoencoder, Cell Reports Medicine, 2024) pretrains by masking and reconstructing marker values, then fine-tunes into an immunophenotyping model that the authors report performs well while retaining biological interpretability. Worth knowing: the public model is built around a specific panel — it expects a defined marker set and order — so it’s a strong self-supervised annotator rather than a panel-agnostic one.
• FATE (Feature-Agnostic Transformer-based Encoder, WACV 2024) is the cleaner example of panel-agnostic design: a set-transformer that doesn’t assume a fixed set of markers, learning embeddings intended to support transfer across samples with different feature sets.
• GPCT (Generalised Pretrained Cytometry Transformer) — emerging, preprint-stage work — targets sample-level prediction across heterogeneous panels via a “universal” per-cell marker embedding, and the authors propose mechanisms intended to highlight which cell subsets contribute to predictions.

These efforts are tackling what is arguably the most important open problem in the field, and it’s the most exciting part of the space right now. The honest caveats are just as real. These models need a lot of data to pretrain; the strongest published ones are often tied to a particular panel, so true cross-panel generalization — exactly what FATE and GPCT are reaching for — is still being established, and robustness across instruments, staining protocols, and laboratories remains much less established than benchmark datasets sometimes suggest. And “interpretable” is a spectrum, not a checkbox: a learned latent space that beats manual gating in a paper can still be a black box you can’t defend line by line during a manuscript revision.

Wave 3: agentic copilots

The frontier is agentic systems — LLM-driven assistants that take a plain-language goal and orchestrate the analysis themselves. It’s worth being precise about where this is actually real: the strongest current examples are in single-cell RNA-seq and broader single-cell omics, not flow cytometry. LLM-based agents for omics analysis were benchmarked head-to-head in 2026, and frameworks like AstraZeneca’s CellAtria brought dialogue-driven, document-to-analysis automation to scRNA-seq. Flow-cytometry-specific agentic analysis is earlier and far less established.

There are defensible reasons flow is behind. In clinical flow cytometry, where results may influence diagnosis or MRD assessment — AI-assisted, human-in-the-loop workflows have been reported in specific leukemia contexts — the bar for handing decisions to an autonomous agent is understandably higher than in exploratory research. And the same data heterogeneity that challenges foundation models challenges agents too.

Our own stance — which we’ve written about before — is copilot, not autopilot: an assistant that proposes gates, computes statistics, and explains its reasoning while a human stays in the loop and owns the final call. An agent that quietly redraws a gate and changes a reported frequency is not a time-saver; it’s a liability.

What we actually trust it for today

Our rule of thumb is boring, and that’s the point:

• Use AI for triage and scale — first-pass gating, batch annotation, surfacing populations you might have missed — not for the final reported result.
• Always validate against manual on a subset. If the automated frequencies don’t track hand-gated ground truth on a representative sample, the speed-up isn’t real, it’s just faster.
• Prefer methods you can interrogate. Given two comparable options, the one that tells you why — which markers, which subsets — wins, because that’s what survives peer review.
• Match the tool to the step. Spectral unmixing, QC, and panel design still reward classic, well-understood methods. Learned models earn their place in annotation and cross-sample integration, where the panel-agnostic work above is most promising.

The exciting part isn’t that AI will replace the immunologist reading the plots. It’s that much of the slow, repetitive work can be compressed so expert judgment goes where it actually matters — which is exactly why we built our own flow tooling around a human-in-the-loop copilot rather than a one-click black box.

If you’re weighing whether to fold any of this into your own flow work — or whether a result from one of these models is solid enough to put in a figure — that’s precisely the kind of call we help labs make through our flow cytometry analysis.