Welcome!

1. Introduction and installation

STEM is a tool for building single-cell level spatial transcriptomic landscapes using SC data with ST data. STEM extracts the spatial information from the gene expressions and eliminates the domain gap between spatial transcriptomics and single-cell RNA-seq data.

Installation

conda install pytorch torchvision torchaudio pytorch-cuda=11.7 -c pytorch -c nvidia ## more info: https://pytorch.org/get-started/locally/
conda install -c conda-forge scanpy python-igraph leidenalg
conda install seaborn

Since STEM is a light model, you can use the code from the STEM folder directly. You can also install STEM from PYPI:pip install scSTEM. To verify the installation, run python test.py

Then we will demonstrate the workflow of generating the results shown in our figure 2.

%pylabinlineimportosos.environ["CUDA_VISIBLE_DEVICES"] ="3"importscanpyasscimportpandasaspdimporttorchimportscipyimporttimefromSTEM.modelimport*fromSTEM.utilsimport*

2. Load and Preprocess data

First, we load the simulated ST data as ST data and the raw SeqFISH data as SC data. Then we normalized and log-scaled these data.

scdata=pd.read_csv('./data/mousedata_2020/E1z2/simu_sc_counts.csv',index_col=0)
scdata=scdata.Tstdata=pd.read_csv('data/mousedata_2020/E1z2/simu_st_counts.csv',index_col=0)
stdata=stdata.Tstgtcelltype=pd.read_csv('./data/mousedata_2020/E1z2/simu_st_celltype.csv',index_col=0)
spcoor=pd.read_csv('./data/mousedata_2020/E1z2/simu_st_metadata.csv',index_col=0)
scmetadata=pd.read_csv('./data/mousedata_2020/E1z2/metadata.csv',index_col=0)
adata=sc.AnnData(scdata,obs=scmetadata)
sc.pp.normalize_total(adata)
sc.pp.log1p(adata)
scdata=pd.DataFrame(adata.X,index=adata.obs_names,columns=adata.var_names)
stadata=sc.AnnData(stdata)
sc.pp.normalize_total(stadata)
sc.pp.log1p(stadata)
stdata=pd.DataFrame(stadata.X,index=stadata.obs_names,columns=stadata.var_names)
adata.obsm['spatial'] =scmetadata[['x_global','y_global']].valuesstadata.obsm['spatial'] =spcoor

Next we calculate the ratio between the median total counts value of SC and ST data as the dropout rate.

sc.pp.calculate_qc_metrics(adata,percent_top=None, log1p=False, inplace=True)
adata.obs['n_genes_by_counts'].median()
sc.pp.calculate_qc_metrics(stadata,percent_top=None, log1p=False, inplace=True)
stadata.obs['n_genes_by_counts'].median()
dp=1-adata.obs['n_genes_by_counts'].median()/stadata.obs['n_genes_by_counts'].median()
#0.5836734693877551

3. Train the STEM model

We first config the STEM model and then train it. Empirically we found by setting the sigma as half of the ST spot adjacent distance, STEM achieves the best performance.

classsetting( object ):
passseed_all(2022)
opt=setting()
setattr(opt, 'device', 'cuda:0') # devicesetattr(opt, 'outf', 'log/test') # folder to save log filessetattr(opt, 'n_genes', 351) # number of genes for the inputsetattr(opt, 'no_bn', False) # duplicatedsetattr(opt, 'lr', 0.002) # learning ratesetattr(opt, 'sigma', 3) # the spatial variance parameter in the Gaussian functionsetattr(opt, 'alpha', 0.8) # MMD loss weight default:0.8setattr(opt, 'verbose', True) # verbosesetattr(opt, 'mmdbatch', 1000) # batch for MMD losssetattr(opt, 'dp', dp) # dropout rate for ST datatestmodel=SOmodel(opt)
testmodel.togpu()
loss_curve=testmodel.train_wholedata(400,torch.tensor(scdata.values).float(),torch.tensor(stdata.values).float(),torch.tensor(spcoor.values).float())

The loss curve will be like this:

4. Get embeddings and reconstruct spatial adjacency

We first get the embeddings and build the mapping matrix We first get the embeddings and build the mapping matrix

testmodel.modeleval()
scembedding=testmodel.netE(torch.tensor(scdata.values,dtype=torch.float32).cuda())
stembedding=testmodel.netE(torch.tensor(stdata.values,dtype=torch.float32).cuda())
netst2sc=F.softmax(stembedding.mm(scembedding.t()),dim=1).detach().cpu().numpy()
netsc2st=F.softmax(scembedding.mm(stembedding.t()),dim=1).detach().cpu().numpy()

The matrix netst2sc and netsc2st are ST-SC and SC-ST mapping matrices, respectively. In the ST-SC mapping matrix, the probability of one spot to other cells summarizes to 1. In the SC-ST mapping matrix, the probability of one cell to other spots summarizes to 1.

Then we can get the spatial coordinate for every single cell.

adata.obsm['spatialDA'] =all_coord(pd.DataFrame(netsc2st,index=adata.obs_names,columns=stadata.obs_names),spcoor)

Compared with other methods, STEM is the only method that preserves the original topology structure of all single cells.

More demos can be found in the Demo folder. Codes for repruducing the results are in the SourceforFigure folder.

Data

The processed data and trained models used for reproducing the results are deposited in Figshare.