search

JinmiaoChenLab / GraphST

GNU Affero General Public License v3.0
91 stars 13 forks source link

Question about learning the mapping matrix for scRNA-seq and ST data integration(equation 9 and 10 in paper) #18

Open UnkownNames opened 11 months ago

UnkownNames commented 11 months ago

Great work, I am interested in this beautiful computational tool for spatial transriptomics. But I have some doubt about the equation (9) and (10) in paper

image image

what is the motivation of using autoencoder to pretrain scRNA-data(just for denoising??), why not using raw data of scRNA-data directly, pretraining autoencoder will improve the deconvolution result significantly? Have you conduct some ablation experiment to explore the influence of pretraining autoencoder? thank you for your suggestions.

longyahui commented 9 months ago

Hi,

Thank you for your questions.

Generally, raw single-cell data is noisy and high dimension. Pretraining scRNA-seq data based on autoencoder will help reduce the noise and dimension, thereby facilitating the improvement of performance in downstream cell type deconvolution task. To validate it, we have conducted ablation studies. Hope it helps. Thank you again.

Best regards,
Yahui