Here we can automatically summarize the overall cellular landscape of a sample profiled with a single-cell technology (e.g. CyTOF, Flow Cytometry, scRNA-seq) with a single feature vector. Such feature vectors can therefore be used in downstream tasks, such as for prediction of clinical or experimental outcomes. Our method, originally introduced in in our 2022 ACM-BCB 2022 paper 'Transparent Single-Cell Set Classification with Kernel Mean Embedding' paper offers an unsupervised approach for summarizing nuances of cellular heterogeneity across multiple profiled samples. Our method is clustering and gating free and simply requires original data matrices of cells x features as inputs. Breiefly, the summarization works by computing a Kernel Mean Embedding on Random Fourier Features.
Dependencies
- Python >= 3.6, anndata 0.7.6, numpy 1.22.4, scipy 1.7.1, tqdm 4.64.0, scanpy 1.8.1
You can clone the git repository by,
git clone https://github.com/CompCy-lab/cellLandscape.git
To compute the kernel mean embedding, first read in a preprocessed .h5ad object. This dataset contains multiple profiled single-cell samples. That is, the .h5ad file contains cells across all samples.
import scanpy as sc
adata = sc.read_h5ad('nk_cell_preprocessed.h5ad')Then simply compute the kernel mean embedding as,
# Inputs:
# adata: annotated data object (dimensions = cells x features)
# sample_set_key: string referring to the key within adata.obs that contains the samples to compute the embedding
# ~ If sample_set_key is None, will use all cells as a single sample
# gamma: scale for standard deviation of the normal distribution within random Fourier frequency feature computation
# D: dimensionality of the random Fourier frequency features, D/2 sin and D/2 cos basis
# frequency_seed: random state parameter
# -----------------------
# Returns:
# X_embed: mean embedding (dimensions = samples x D)
# -----------------------
from embed import mean_embed
X_embed = mean_embed(adata, sample_set_key = 'FCS_File', gamma = 1, D = 2000, frequency_seed = 0)