abuchin/compass

Vector representations of gene co-expression in single cell RNAseq.

compass 🧭

Vector representations of gene co-expression in single cell RNAseq.

Install:

python3 -m venv cenv
python3 setup.py install

Loading scanpy datasets into Compass.

from compass.data import Context, CompassDataset
context = Context.build(adata)
dataset = CompassDataset(context)

Training gene vectors.

from compass.model import CompassTrainer
cmps = CompassTrainer(dataset,output_file="genes.vec", batch_size=200)
cmps.train(10) # run for 10 iterations

Loading results.

from compass.embedding import GeneEmbedding, CellEmbedding
gembed = GeneEmbedding("genes.vec", context)
cembed = CellEmbedding(context, gembed)

Compute gene similarities.

gembed.compute_similarities("CD8A")

Clustering gene vectors.

gene_clusters = gembed.cluster()

Build gene t-SNE embedding and label some genes.

gembed.plot(gene_clusters,labels=["C1QC","C1QA","TYROBP"])

Find most representative genes for each cluster based on cosine distance.

cluster_definitions = gembed.cluster_definitions(gene_clusters)

Pandas data frame of most representative genes for each cluster.

gene_df = gembed.cluster_definitions_as_df(cluster_definitions,top_n=10)

Relabel cluster in results.

gembed.relabel_cluster(cluster_definitions, gene_clusters, 6, "Cell Cycle")

Generate an average vector for a set of genes.

cd8tcellvec = gembed.generate_vector(["CD8A","CD8B","CD3D","CD3E","CD3G"])

Plot a cosine similarity matrix for a set of genes.

gembed.plot_similarity_matrix(["CD8A","CD8B","CD3D","CD3E","CD3G"], png="cd8_matrix.png")

Plot a networkx graph of similarities between a set of genes.

plot_similarity_network(["CD8A","CD8B","CD3D","CD3E","CD3G"], png="cd8_graph.png"

Build cell t-SNE embedding with cell type labels.

tsne_embedding = cembed.plot(column="cell_type")

Reuse t-SNE embedding and plot another label.

cembed.plot(column="batch_label",pcs=tsne_embedding)

Find K cell clusters.

clusters = cembed.cluster(k=6)

Locally batch correct on clusters using column as batch label.

cembed.batch_correct(column="batch_label",clusters=clusters)

Regenerate corrected t-SNE

corrected_tsne_embedding = cembed.plot(column="batchlb")

Plot cosine distance for each cell to a given vector.

cembed.plot_distance(cd8tcellvec,png="highlightcd8cells.png")