flowkit – Data & Science skill for AI agents

FlowKit

When to use

Use FlowKit to apply or build cytometry gating strategies, analyze batches of

FCS samples, or reproduce supported FlowJo workspace analyses. It supports

GatingML 2.0 and a subset of FlowJo 10 features. Import success alone does

not establish agreement with FlowJo.

The examples and bundled helper target FlowKit 1.3.2 on Python 3.13.

Upstream supports additional Python versions; those were not exercised here.

The helper and examples were tested on synthetic FCS data, including a public

FlowJo 10.7.1 synthetic workspace fixture. They are not biological validation.

Install

Use a separate environment; FlowKit 1.3.2 requires NumPy >2 and pandas <3:

uv venv --python 3.13 .venv-flowkit
uv pip install --python .venv-flowkit/bin/python "flowkit==1.3.2"
.venv-flowkit/bin/python -c "import flowkit; print(flowkit.__version__)"

The scientific package is BSD-3-Clause licensed; this skill is MIT licensed.

Workflow

1. Identify the analysis definition. Use Session for a programmatic or

GatingML strategy; use Workspace for FlowJo sample-specific gates,

compensation, and transforms. Request the actual strategy or controls when

biological thresholds have not been supplied.

2. Inspect samples and channel identities. Match detector/PnN labels to

compensation matrices and gate dimensions; PnS marker names may be empty or

repeated. Verify sample IDs: the default is FCS $FIL, which can differ

from the current filename. Reject ID collisions before loading a batch.

3. Establish the coordinate system. Determine whether the supplied events

are already compensated. Apply compensation before nonlinear transforms;

match gate thresholds to the same transformed or untransformed coordinates.

See [compensation and gating](references/compensation-and-gating.md).

4. Check the hierarchy. Preserve parent gates and full gate paths, including

root. For a study, review acquisition/time stability, debris exclusion,

singlets, viability, and phenotype gates as appropriate to its panel. Use

single-stain controls for compensation and suitable negative/FMO controls

for positivity; demonstration thresholds are not transferable biology.

5. Analyze and inspect. Run on all events, then check gate overlays and

sample-level QC. A plot's subsample is not the population denominator.

Review warnings and compare representative imported results to FlowJo.

6. Export counts with denominators and provenance. Keep gate paths,

sample IDs, total event counts, input hashes, package versions, and the

analysis definition. Keep biological replicates identifiable; events from

one specimen are not independent experimental replicates.

Apply an existing strategy

Set FLOWKIT_SKILL_DIR to this skill's installed directory. From the repository

root it is skills/flowkit. Paths below represent the user's local inputs.

FLOWKIT_SKILL_DIR="skills/flowkit"
uv run --no-project --python 3.13 --with "flowkit==1.3.2" \
  python "$FLOWKIT_SKILL_DIR/scripts/analyze_gates.py" \
  --gatingml gates.xml --fcs sample.fcs --output-dir results-gatingml

For a FlowJo workspace, supply every FCS file in the selected group:

uv run --no-project --python 3.13 --with "flowkit==1.3.2" \
  python "$FLOWKIT_SKILL_DIR/scripts/analyze_gates.py" \
  --workspace study.wsp --group "Study" \
  --fcs sample-a.fcs sample-b.fcs --output-dir results-workspace

The helper writes gate_report.csv and provenance.json to a new directory.

Each row includes sample_event_count, parent_event_count, and a full

population_path; empty-parent percentages are blank and flagged with

relative_percent_defined=False.

It rejects duplicate sample IDs, missing/extra workspace-group samples,

zero-event samples, and strategies without gates. It uses explicit input files,

does not follow paths embedded in the workspace, and runs without

multiprocessing or transformed-event caching. It still loads each sample into

memory; use manageable batches via the Python API for large studies.

--filename-as-id deliberately switches from $FIL to file basenames. Use it

only when those names match the analysis definition. See

[workspace analysis](references/workspaces-and-results.md) for partial-group

analysis, result interpretation, and fluorescence summaries.

Build a strategy in Python

This runnable example uses sample.fcs with FSC-A, FL1-A, and FL2-A.

The matrix, thresholds, and transform parameters are **synthetic teaching

values**. Replace them with the study's validated settings.

import flowkit as fk
import numpy as np

sample = fk.Sample("sample.fcs")
strategy = fk.GatingStrategy()
strategy.add_comp_matrix(
    "spill", fk.Matrix(
        np.array([[1.0, 0.1], [0.2, 1.0]]), ["FL1-A", "FL2-A"],
        fluorochromes=["FITC", "PE"],
    )
)
logicle = fk.transforms.LogicleTransform(
    param_t=262144, param_w=0.5, param_m=4.5, param_a=0
)
strategy.add_transform("logicle", logicle)
strategy.add_gate(
    fk.gates.RectangleGate("Cells", [
        fk.Dimension("FSC-A", range_min=50, range_max=300)
    ]),
    gate_path=("root",),
)
thresholds = logicle.apply(np.array([50.0, 600.0]))
strategy.add_gate(
    fk.gates.RectangleGate("Positive", [
        fk.Dimension(
            "FL1-A", compensation_ref="spill", transformation_ref="logicle",
            range_min=float(thresholds[0]), range_max=float(thresholds[1]),
        )
    ]),
    gate_path=("root", "Cells"),
)
session = fk.Session(gating_strategy=strategy, fcs_samples=[sample])
session.analyze_samples(use_mp=False)
report = session.get_analysis_report()
print(report[["sample_id", "gate_path", "gate_name", "count",
              "absolute_percent", "relative_percent"]])
with open("gates.xml", "xb") as handle:
    session.export_gml(handle)

GatingML exports a template by default. When custom per-sample gates exist,

use session.export_gml(handle, sample_id=sample.id) for that sample's strategy.

A single template export does not preserve every sample-specific override.

Interpretation checks

• count is the number of events passing the gate and its ancestors.
• absolute_percent is percent of all sample events; relative_percent is

percent of the immediate parent. These are percentages, not fractions.

• Gate names can repeat under different parents. In the helper's output use

(sample_id, population_path) as the identifier. Paths are JSON arrays

inside CSV cells. FlowKit's native report stores a quadrant's owner

separately in quadrant_parent; its gate_path alone omits that owner.

• A zero-event parent makes a child percentage biologically undefined;

FlowKit 1.3.2 reports zero for ordinary children and NaN for quadrants.

The helper exports both as blank with an explicit false flag. This differs

from a defined 0% for an empty gate whose parent contains events.

• Compensated negative fluorescence is legitimate. Do not clip it to zero or

discard those events merely to permit a logarithmic transform.

• Define whether “MFI” means mean or median and name the event source.

A transformed display value is not an intensity on the original scale.

References

• [Compensation and gating](references/compensation-and-gating.md): event

sources, detector order, transform semantics, gate paths, and plotting.

• [Workspaces and results](references/workspaces-and-results.md): sample

matching, missing FCS files, FlowJo limits, and gated fluorescence summaries.

• [FlowKit API](https://flowkit.readthedocs.io/en/latest/api.html) and

[versioned source](https://github.com/whitews/FlowKit/tree/1.3.2): consult

signatures when moving beyond the tested release.

• [FlowKit publication](https://doi.org/10.3389/fimmu.2021.768541): cite the

software and version in scientific methods when used for analysis.