RSTG: Robust Generation of High Quality Spatial Transcriptomics Data using Beta Divergence Based AutoEncoder.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Spatial transcriptomics analysis faces challenges due to limited training data.
• Existing generative models struggle with noisy data, including outliers.
• Developing robust methods for synthetic data generation is crucial for advancing spatial transcriptomics.

Purpose of the Study:

• To propose RSTG (Robust Spatial Transcriptomic Generator), a novel autoencoder model.
• To enhance the generation of realistic and high-quality spatial transcriptomic sequences.
• To improve the robustness of generative models against data noise and outliers.

Main Methods:

• Developed RSTG, an autoencoder incorporating the beta-ELBO loss.
• Utilized variational inference to approximate data distribution and density estimation.
• Validated the model on diverse spatial transcriptomics datasets (MERFISH, MERSCOPE, Visium).

Main Results:

• RSTG demonstrated improved performance in generating high-quality spatial transcriptomic data.
• The model effectively recovered cellular positions in 2D spatial and location domains.
• RSTG maintained data quality and stability even with contaminated training data (outliers, batch effects, dropouts).

Conclusions:

• RSTG offers a robust solution for generating synthetic spatial transcriptomics data.
• The model's ability to handle noise enhances its applicability in real-world scenarios.
• RSTG advances spatial transcriptomics analysis by providing reliable synthetic data for model training.