Canvas to Chat
Select data points on plots or tables and ask Lobster about them.
Canvas to Chat lets you select data directly on your visualizations and ask Lobster about it. Select a cluster of cells on a UMAP, highlight rows in a data table, or pick residues on a protein structure -- Lobster sees exactly what you selected and can analyze it in context.
How It Works
Select on the canvas
Use your mouse to select data points on a plot (box select or lasso), highlight rows in a data table, or click residues on a 3D structure.
Review the prompt
A selection prompt appears near the chat showing what you selected -- for example, "142 cells selected in UMAP (leiden cluster 3)." You can edit or refine it before sending.
Ask Lobster
Click Ask Lobster or press Enter. Lobster receives the full selection context -- which points, from which plot, with what metadata -- and responds with relevant analysis.
Supported Viewers
| Viewer | What you can select | Example question |
|---|---|---|
| Scatter plots | Points via box select or lasso | "What genes are differentially expressed in this cluster?" |
| Data tables | Rows via click or shift-click | "Summarize the metadata for these 50 samples." |
| Protein structures | Residues via click | "What is the function of this binding site?" |
| Pathway maps | Genes and compounds via click | "What role does this enzyme play in the pathway?" |
| Pangenome viewer | Genes and strains | "Compare these selected strains." |
Plot and table selections can reference up to 50,000 data points. Lobster materializes the full selection server-side, so even large selections are analyzed accurately.
Overview
Behind the scenes, each viewer builds a narrative -- a human-readable description of the selection that Lobster reads directly. For plots, this includes axis titles, color grouping, and sample IDs. For tables, it includes the sort/filter context and observation names. This narrative-first design means Lobster understands not just what you selected, but where and why it stands out.
Selections use stable identifiers (cell barcodes for tables, global indices for plots) so they survive sorting and filtering. Selecting 50 cells, re-sorting the table, and then asking a question still references the correct cells.
Tips
- Box select for clusters, lasso for irregular shapes. On scatter plots, use box select (click-drag) for rectangular regions or lasso (shift-drag) for freeform shapes.
- Combine selection with follow-up questions. After Lobster analyzes a selection, ask follow-ups like "Now compare these to the rest of the dataset."
- Selections are per-session. Switching sessions clears the active selection. Each session tracks its own canvas context independently.