desc_paul.md

The Tutorial of Paul et al (2015) for desc

• The reproducible of DESC for Paul's data

import os              
os.environ['PYTHONHASHSEED'] = '0'
import desc          
import pandas as pd                                                    
import numpy as np                                                     
import scanpy.api assc                                                                                 
from time import time                                                       
import sys
import matplotlib
import matplotlib.pyplot as plt
%matplotlib inline 
sc.settings.set_figure_params(dpi=300)

Using TensorFlow backend.

print(sys.version)

3.5.3 |Continuum Analytics, Inc.| (default, Mar 6 2017, 11:58:13) [GCC 4.4.7 20120313 (Red Hat 4.4.7-1)]

sc.logging.print_versions()

scanpy==1.3.6 anndata==0.6.18 numpy==1.14.6 scipy==1.1.0 pandas==0.23.0 scikit-learn==0.19.2 statsmodels==0.9.0 python-igraph==0.7.1 louvain==0.6.1

desc.__version__

'1.0.0.2'

#we have downloaded this data, if not, the following will download data automatically
adata=sc.datasets.paul15()
adata.obs['celltype']=adata.obs['paul15_clusters'].str.split("[0-9]{1,2}", n = 1, expand = True).values[:,1]
adata.obs['celltype2']=adata.obs['paul15_clusters']

WARNING: In Scanpy 0.*, this returned logarithmized data. Now it returns non-logarithmized data. ... storing 'paul15_clusters' as categorical

import desc

sc.pp.log1p(adata)
sc.pp.filter_genes_dispersion(adata,n_top_genes=1000)
sc.pp.scale(adata,max_value=6)
save_dir="paul_result"
adata=desc.train(adata,
        dims=[adata.shape[1],64,32],
        tol=0.005,
        n_neighbors=10,
        batch_size=256,
        louvain_resolution=[0.8,1.0],# not necessarily a list, you can only set one value, like, louvain_resolution=1.0
        save_dir=str(save_dir),
        do_tsne=True,
        learning_rate=200, # the parameter of tsne
        use_GPU=False,
        num_Cores=1, #for reproducible, only use 1 cpu
        num_Cores_tsne=4,
        save_encoder_weights=False,
        save_encoder_step=3,# save_encoder_weights is False, this parameter is not used
        use_ae_weights=False,
        do_umap=False) #if do_uamp is False, it will don't compute umap coordiate

#After training from desc, the results have been saved in `save_dir`
#adata=sc.read("paul_result/adata_desc.h5ad")
adata

AnnData object with n_obs × n_vars = 2730 × 1000 obs: 'paul15_clusters', 'celltype', 'celltype2', 'desc_0.8', 'desc_1.0' var: 'means', 'dispersions', 'dispersions_norm' uns: 'prob_matrix1.0', 'prob_matrix0.8', 'iroot' obsm: 'X_Embeded_z0.8', 'X_tsne', 'X_tsne0.8', 'X_Embeded_z1.0', 'X_tsne1.0'

1. The meta.data of each cell has been saved in adata.obs,
2. the representation from desc of each cell have been saved in adata.obsm('X_Embeded_z1.0')
3. The dimension reduction from desc of each cell have beed saved in adata.obsm('X_tsne1.0')

Computing maxmum probability

adata.obs['max.prob']=adata.uns["prob_matrix1.0"].max(1)

t-SNE plots

sc.pl.scatter(adata,basis="tsne1.0",color=['desc_1.0','max.prob'])

sc.pl.scatter(adata,basis="tsne1.0",color=['celltype','celltype2'])

reference

1. Paul F, Arkin Y, Giladi A, Jaitin DA et al. Transcriptional heterogeneity and lineage commitment in myeloid progenitors. Cell, 163:1663-167 (2015)