search

xmuyulab / DAISM-XMBD

Highly accurate cell type proportion estimation with in silico data augmentation and deep neural networks
MIT License
16 stars 2 forks source link

Question about ground truth proportion #3

Open Chuang1118 opened 3 years ago

Chuang1118 commented 3 years ago

Hello author,

Congratulation you got the competition deconvolution.

I just have a question naive. For DAISM-DNN user must be have califra.txt file ?

I.e. I have scRNAseq(.h5ad) as reference with the annotations and bulk expression file, I want to predict cell type proportion within bulk seq dataset.

DAISM-DNN can handle it ?

Best, Chuang

ZoeLin1130 commented 3 years ago

Hello @Chuang1118 ,

I am sorry for the late reply. You can use 'Generic_simulation' module to generate training dataset if you do not have calibration samples.

daism Generic_simulation -platform S -aug scRNA.h5ad -N 16000 -testexp testexp.txt -outdir ./

And then use Generic_mixsam.txt and Generic_mixfra.txt to train DAISM-DNN.

daism training -trainexp ./output/Generic_mixsam.txt -trainfra ./output/Generic_mixfra.txt -net coarse -outdir ./

Best, Zoe

Chuang1118 commented 3 years ago

Hello, Thanks, For the moment, I always wait for my GPU, since the graphics card is out of stock. How much gpu memory required for DAISM-DNN, if I have about 100,000 single cells as reference and 50 samples bulkseq? I can use the Slurm Cluster with GPU, but it isn't very easy to set up, exceptional for exploiting new API. Do you have singularity images version 3 ?

For example, I have singularity3 image, I can set up my script call GPU mode inside python.

# this line forces theano to use the GPU
os.environ["THEANO_FLAGS"] = 'device=cuda,floatX=' + data_type + ',force_device=True'

I don't know how to work with DAISM-DNN.

Best, Chuang