SURF: A Self-Supervised Deep Learning Method for Reference-Free Deconvolution in Spatial Transcriptomics

Area of Science:

• Genomics
• Computational Biology
• Molecular Biology

Background:

• Spatial transcriptomics offers spatially resolved gene expression but struggles with cellular-level analysis due to signal consolidation in spots.
• Reference-based deconvolution methods often lack necessary matched single-cell data.

Purpose of the Study:

• To develop a reference-free deconvolution tool, SURF, to overcome limitations in spatial transcriptomics.
• To accurately model cellular composition and gene interactions within tissue microenvironments.

Main Methods:

• SURF integrates high-dimensional gene data analysis with self-supervised deep learning.
• It models nonlinear gene interactions and leverages spatial relationships between spots.
• The method was benchmarked on synthetic and real-world spatial transcriptomic datasets.

Main Results:

• SURF consistently outperforms existing reference-free deconvolution methods.
• It achieves performance exceeding reference-based methods when references are unavailable.
• SURF accurately represents diverse tissue microenvironments across various resolutions and species.

Conclusions:

• SURF provides a robust and accurate solution for reference-free deconvolution in spatial transcriptomics.
• The tool precisely models tissue microenvironments and identifies critical biological mechanisms, such as epithelial-to-mesenchymal transition in colorectal cancer metastasis.