Cell embeddings

With hundreds to thousands of cells, the question becomes which cells are alike. uchrom.emb embeds cells by their contact maps — FastHigashi’s tensor decomposition, or scHiCluster’s imputed maps — and by the other measurements of the same cells (RNA, accessibility), clusters them, scores the clusters against known cell types and stores everything on the cells axis (cd.cellm, with the metadata the web browser uses to label it in cd.uns["embeddings"]).

import uchrom as uc
import uchrom.datasets as ds
import uchrom.emb as emb

cd = ds.load("kim2020_scihic")                          # cells + per-cell contact maps (linked)
uc.tl.embed_higashi(cd, rank=64, seed=0)                # FastHigashi -> cd.cellm["higashi"]
emb.score_embedding(cd, "higashi", "cell_type")         # ARI / NMI of k-means, kNN purity

Task	API
embed cells by their contact maps: FastHigashi	uc.tl.embed_higashi / uchrom.emb.embed_higashi (extra emb)
embed cells by their contact maps: scHiCluster	uc.tl.embed_contacts / uchrom.emb.embed_contacts
embed cells from RNA / ATAC / chromatin marks / any matrix (PCA, t-SNE, UMAP)	uc.tl.embed_cells / uchrom.emb.embed_cells; RNA from the linked AnnData: add_gene_counts (highly variable genes)
clusters and their markers	uc.tl.cluster_cells, marker_features
how well an embedding separates known types	score_embedding, cluster_agreement, knn_purity
plot an embedding, coloured by a type or a value	plot_embedding

The same functions embed imaging cells by their chromatin marks (DNA seqFISH+).

Tutorials

• Contact-map cell embedding with FastHigashi
• Cell embeddings and clusters across modalities

API: uchrom.emb.