Improving cell-type composition inference in spatial transcriptomics with SpaDAMA.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Accurate cell-type composition is crucial for disease research and understanding tissue heterogeneity.
• Current spatial transcriptomics (ST) methods lack single-cell resolution, hindering precise cell identification.
• Existing deconvolution methods using single-cell RNA sequencing (scRNA-seq) data often fail to account for domain-specific data differences.

Purpose of the Study:

• To develop a novel computational method for accurate cell-type deconvolution in spatial transcriptomics data.
• To address the challenge of data discrepancies between scRNA-seq and ST data modalities.
• To enhance the identification of cellular components within complex tissue microenvironments.

Main Methods:

• Introduced Domain-Adversarial Masked Autoencoder (SpaDAMA), a novel method for cell-type deconvolution.
• Utilized Domain-Adversarial Learning (DAL) to harmonize scRNA-seq and ST data distributions.
• Implemented masking strategies to enhance feature extraction and mitigate noise/artifacts in ST data.

Main Results:

• SpaDAMA demonstrated superior performance in cell-type deconvolution across 32 simulated and 4 real-world datasets.
• The method effectively harmonized data from different modalities, creating a unified latent representation.
• SpaDAMA successfully minimized discrepancies and noise, leading to more reliable cell-type composition estimation.

Conclusions:

• SpaDAMA offers a robust and accurate solution for cell-type deconvolution in spatial transcriptomics.
• The method provides a valuable tool for advancing disease target identification and tissue heterogeneity studies.
• SpaDAMA's domain-adversarial approach effectively bridges the gap between scRNA-seq and ST data.