omicverse.pp.tsne

Contents

omicverse.pp.tsne#

omicverse.pp.tsne(adata, n_pcs=None, *, n_components=2, use_rep=None, perplexity=30, metric='euclidean', early_exaggeration=12, learning_rate=1000, random_state=0, key_added=None, copy=False, n_iter=None, **kwargs)[source]#

Compute t-SNE coordinates for cells, dispatching by ov.settings.mode.

Parameters:

  • adata (anndata.AnnData) – AnnData with PCA (or another use_rep) available.

  • n_pcs (default: None) – Number of PCs to use from adata.obsm['X_pca'].

  • n_components (int (default: 2)) – Output dimensions (typically 2).

  • use_rep (str | None (default: None)) – .obsm key to compute t-SNE on; defaults to X_pca.

  • perplexity (float (default: 30)) – Controls the effective number of nearest neighbors used to construct the high-D affinities. Typical range 5–50.

  • metric (str (default: 'euclidean')) – Distance metric for neighbor search. 'euclidean' by default.

  • early_exaggeration (float (default: 12)) – Scaling applied to P during the early-exaggeration phase (first ~250 iterations). Larger makes clusters tighter in the output.

  • learning_rate (float (default: 1000)) – Learning rate of the optimiser. 1000 is the sklearn default; the GPU backend auto-rescales to a much smaller value internally.

  • random_state (default: 0) – Seed for reproducibility.

  • key_added (str | None (default: None)) – Target .obsm / .uns key. Defaults to 'X_tsne' / 'tsne'.

  • copy (bool (default: False)) – Return a new AnnData instead of modifying in place.

  • n_iter (int | None (default: None)) – Total optimisation iterations. None uses the backend default.

  • **kwargs – Forwarded to the backend (scanpy.tl.tsne, omicverse.pp._tsne.tsne, or rapids_singlecell.tl.tsne).

Notes

cpu-gpu-mixed mode uses our KL-loss native torch t-SNE (see omicverse/external/torch_tsne.py) — no torchdr dependency.

Contents