We propose SpaDAMA, a Domain-Adversarial Masked Autoencoder, for cell-type deconvolution in spatial transcriptomics. Utilizing Domain-Adversarial learning, SpaDAMA aligns real ST data with simulated ST data derived from scRNA-seq, mapping both into a common latent space to reduce the modality gap. Additionally, it employs masking strategies to strengthen feature learning from real ST data while suppressing noise and spatial confounders. (Variational)

To run SpaDAMA, you need to install PyTorch with GPU support first. The environment supporting SpaDAMA and baseline models is specified in the requirements.txt file.

The publicly available datasets were used in this study. You can download them from https://doi.org/10.5281/zenodo.14221635

After configuring the environment, download dataset4 from the Simulated_datasets in the data repository and place it into the Simulated_datasets folder. Then, Run main_code.pyto start the process.If you want to run other data, simply modify the file path.

SpaDAMA: Spatial Transcriptomics Deconvolution Using Domain-Adversarial Masked Autoencoder

If you have any questions, please contact huanglin212@aliyun.com and minwenwen@ynu.edu.cn