Attention-guided enhanced deconvolution enables reference-free cell type estimation in spatial transcriptomics.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Spatial transcriptomics provides gene expression data with spatial context.
• Current deconvolution methods often require single-cell atlases or manual parameter tuning.
• Challenges exist in accurately identifying cellular composition from mixed signals in spatial transcriptomics.

Purpose of the Study:

• To develop a reference-free computational framework for deconvolution of spatial transcriptomics data.
• To overcome limitations of existing deconvolution approaches, particularly the reliance on single-cell references.
• To enable accurate cell type identification and spatial mapping in diverse biological tissues.

Main Methods:

• Developed Attention-Guided Enhanced Deconvolution (AGED), a two-stage framework combining probabilistic modeling and neural attention.
• Stage 1: Performer-based network with linear-complexity attention for automatic selection of optimal cell type numbers.
• Stage 2: Attention-Guide refines cell type features using cross-attention, spatial attention, and collaborative attention with dynamic gating.

Main Results:

• AGED automatically identified anatomical structures in Mouse Olfactory Bulb (MOB) tissue with high reconstruction performance (r=0.86).
• The method revealed detailed cell type distributions and relationships in human pancreatic ductal adenocarcinoma (PDAC) and thymus tissues.
• Learned cell type distributions demonstrated biological interpretability, aligning with known anatomical boundaries and molecular markers.

Conclusions:

• AGED offers a practical and effective reference-free solution for spatial transcriptomics deconvolution.
• The framework accurately identifies cellular composition and maintains biological interpretability across diverse tissues.
• This approach advances spatial transcriptomics analysis by removing the need for matched single-cell data.