FlashDeconv reveals resolution horizons in atlas-scale spatial transcriptomics.

Area of Science:

• Computational Biology
• Genomics
• Bioinformatics

Background:

• Atlas-scale spatial transcriptomics demands efficient deconvolution methods that preserve rare biological signals.
• Existing methods often struggle with computational costs or conflate biological information with population abundance.
• Rare cell type signals are frequently lost during standard feature selection processes.

Purpose of the Study:

• To introduce FlashDeconv, a novel framework for scalable and computationally efficient spatial transcriptomics deconvolution.
• To preserve rare cell type signals often discarded by conventional feature selection methods.
• To enable rapid analysis of large-scale spatial transcriptomics data for biological discovery and clinical applications.

Main Methods:

• FlashDeconv utilizes structure-preserving randomized sketching and leverage-score importance sampling.
• This approach prioritizes transcriptomically distinct markers to preserve rare cell signals.
• The framework enables systematic scale-space exploration by defining a 'resolution horizon'.

Main Results:

• FlashDeconv demonstrates accuracy comparable to leading Bayesian methods but accelerates inference by orders of magnitude.
• Applied to human ovarian cancer cohorts, it rapidly reproduces clinical response signatures, enabling quick patient stratification.
• Operating below the 'resolution horizon' (8-16 µm), FlashDeconv uncovers cryptic Tuft cell niches enriched for intestinal stem cells.

Conclusions:

• FlashDeconv offers a scalable, mathematically grounded framework for atlas-scale spatial transcriptomics discovery.
• The method overcomes limitations of variance-based approaches and coarse spatial binning.
• It provides a powerful tool for uncovering complex biological architecture and enabling rapid clinical insights.