AI Features
How AI gating works
Cytomaton uses a per-user k-nearest neighbor (k-NN) model that learns from your own gate history. It matches the current panel type (based on parameter names) against panels you've gated before and suggests gate boundaries based on your prior analyses.
The AI gets better the more you use it — and it never trains on anyone else's data. Your FCS data is never used to train external models.
When AI gating activates
AI suggestions appear after you've drawn 3 gates on the same panel type (same set of parameter names). Until then, the workspace shows a progress indicator: "N more gates to activate AI."
Recognizing AI suggestions
AI-suggested gates appear as dashed overlays on the scatter plot with a confidence badge inside the overlay. Hovering over a gate in the gating tree shows an AI provenance tooltip: "AI suggested — based on X prior analyses."
Confidence levels:
| Confidence | Meaning |
|---|---|
| ≥75% | High confidence. "Accept all ≥75%" button applies these automatically |
| 60–74% | Moderate. Review individually before accepting |
| <60% | Low. Suggestions are suppressed (not shown) |
Accept-and-adjust workflow
Click Accept on a single suggestion to place it in a pending state, then drag any vertex to refine the boundaries. Click Confirm to finalize. This lets you use the AI's starting point while maintaining full control over the final gate.
AI provenance in exports
When you enable AI provenance export, a methods text paragraph is included with publication figures stating that AI suggestions were used and the confidence level. This is optional and controlled per export — useful for transparent reporting in publications.
AI chat assistant
The AI chat panel is always visible in the app sidebar. It's a domain-expert flow cytometry assistant that knows about your loaded FCS files, current gates, compensation/unmixing status, and QC flags.
What the AI chat can help with: - Gating strategies for specific panel types - Compensation and unmixing guidance - Panel design recommendations - Data interpretation and troubleshooting - FlowSOM and dimensionality reduction guidance
Model selection: Claude Sonnet 4.6, Claude Haiku 4.5, GPT-4o, or GPT-4o Mini. You can also bring your own API key — keys are encrypted server-side and never exposed in the browser.
Important: AI responses are for guidance only. Verify all analytical decisions independently. The AI never generates gate coordinates — gating decisions are always the researcher's.