Node reference

Annotation nodes.

Cell type annotation by four complementary routes. Auto Annotation asks an LLM (OpenAI / Anthropic / Gemini, your key) to name each cluster from its marker genes; CellTypist applies a pretrained logistic-regression model per cell; Cell Mapper transfers labels from a reference atlas by kNN on a joint PCA; Annotation Concordance compares any two or more label columns (ARI / NMI, cross-tab, agreement matrix) so you can judge which labels to trust.

04 nodesCategory · annotationin 2 templates

Heads up

CellTypist models are installed through the Resource Center before a run — the node never downloads one itself. A missing model stops the node with a message naming the resource.

Auto Annotation sends cluster marker lists (gene symbols and statistics) to the LLM provider you chose, never expression matrices. It is not deterministic; record the model and the result in your project.

Auto Annotation

autoAnnotation

LLM-based cell type annotation (OpenAI / Anthropic / Gemini). No reference dataset needed. Computes cluster markers, sends to LLM with tissue context, predicts cell types. Cost ~$0.01-0.30 per run depending on model.

Node LibraryscRNA-seqSpatial

InAnnData (with clustering)

OutAnnData (obs["cell_type_predicted"], obs["cell_type_predicted_confidence"])

Parameters (10)

Name	Type	Default	Description
`species`	select	`human`	human/mouse/rat/zebrafish/drosophila/c.elegans/other.
`tissue`	text	—	Tissue (e.g. "PBMC", "brain cortex"). Required.
`cluster_key`	obs column	`leiden`	Cluster obs column.
`provider`	select	`openai`	openai/anthropic/gemini.
`model`	select	`gpt-5.6-terra`	Model name.
`use_chat_key`	checkbox	`true`	Use API key from AI Chat Settings.
`api_key`	text	—	Override API key for this node only. Only shown when 'Use Chat API Key' is unchecked.
`n_top_markers`	number	`7`	Top markers per cluster sent to LLM.
`restrict_to_expected`	checkbox	`false`	Only allow LLM-listed cell types.
`result_key`	text	`cell_type_predicted`	Output obs column.

Cell Mapper

cellMapper

Reference-based label transfer via kNN mapping on joint PCA. Computes joint PCA on common genes, finds nearest neighbors in reference, transfers labels via kernel-weighted voting.

Node LibraryscRNA-seqSpatial

InAnnData (preprocessed) + reference .h5ad

OutAnnData (obs["{key}_transferred"], optional obsm["X_umap_transferred"])

Parameters (7)

Name	Type	Default	Description
`reference_file`	file (.h5ad)	—	Reference atlas with annotated cell types.
`obs_keys`	text	`celltype`	Comma-separated obs columns to transfer.
`use_rep`	select	`auto`	"auto" joint PCA, or shared embedding key.
`n_neighbors`	number	`15`	Reference neighbors per query cell.
`kernel`	select	`jaccard`	jaccard/gauss/inverse_distance/equal.
`transfer_umap`	checkbox	`false`	Transfer UMAP coordinates.
`result_suffix`	text	`_transferred`	Suffix for transferred columns.

CellTypist Annotation

celltypistAnnotation

Per-cell type annotation using CellTypist pretrained logistic-regression models — no LLM and no custom reference required. Applies a downloadable pretrained CellTypist model (e.g. Immune_All_Low.pkl) to label every cell directly. Optional majority voting harmonises predictions within over-clustered neighbourhoods. Requires log1p-normalised (to 10,000 counts) expression in adata.X.

Node LibraryscRNA-seqSpatial

InAnnData (log1p-normalised; run Normalize + Log1p upstream)

OutAnnData (obs["celltypist_labels"] + obs["celltypist_labels_conf"])

Parameters (6)

Name	Type	Default	Description
`model`	text	`Immune_All_Low.pkl`	CellTypist model name or path to a .pkl. Download models in the Model Setup panel.
`majority_voting`	checkbox	`true`	Refine per-cell predictions by majority vote within over-clustered neighbourhoods.
`over_clustering`	obs column	—	obs column for majority voting. Empty = CellTypist over-clusters automatically.
`mode`	select	`best match`	"best match" = single top label; "prob match" = multi-label / Unassigned by threshold.
`p_thres`	number	`0.5`	Probability cutoff used in "prob match" mode.
`result_key`	text	`celltypist_labels`	obs column for labels. Confidence saved to <result_key>_conf.

Annotation Concordance

annotationConcordance

Compare 2+ categorical annotation/cluster columns and visualise their agreement — a QC step to judge whether an automatic annotation can be trusted. Two plot types: "crosstab" (row-normalised confusion heatmap of two label columns) and "agreement_matrix" (pairwise ARI or NMI across all label columns). High agreement between methods like CellTypist, Auto Annotation, Cell Mapper, and Leiden means the labels are reliable. All tuning lives in the Visualization Workspace.

Node LibraryscRNA-seq

InAnnData (2+ categorical obs columns)

OutPNG image (agreement_matrix also returns ARI/NMI scores)

Plots (2)crosstabagreement_matrix

Parameters (13)

Name	Type	Default	Description
`plot_type`	text	`crosstab`	"crosstab" = confusion heatmap of the first two; "agreement_matrix" = ARI/NMI across all.
`label_keys`	text	—	Two or more comma-separated obs columns. Empty = auto-detect categorical columns.
`metric`	select	`ari`	agreement_matrix only: "ari" or "nmi".
`normalize`	select	`index`	crosstab only: "index" (rows sum to 1), "columns", or "none".
`annotate`	checkbox	`true`	Show numeric values in heatmap cells.
`cmap`	select	`viridis`	Colormap.
`value_fmt`	select	—	Number format for the in-cell annotation (e.g. '.2f', 'd'). Empty = auto (0 decimals for raw counts, 2 for normalized).
`mask_diagonal`	checkbox	`false`	[agreement_matrix] Hide the self-comparison diagonal (always 1.0) to rescale the colour range to the off-diagonal agreement.
`min_count`	number	—	[crosstab] Hide cells with fewer than this many cells. Empty (null) = show all.
`figsize_w`	number	`10`	Figure width in inches.
`figsize_h`	number	`8`	Figure height in inches.
`font_size`	number	—	Base font size. Empty = matplotlib default.
`publication_quality`	checkbox	`false`	Render at 300 DPI for publication figures.

CellRank

Pathway