Multi-scale spatial testing recovers gene programs missed by existing detection methods.

Area of Science:

• Genomics
• Computational Biology
• Bioinformatics

Background:

• Accurate and scalable methods are crucial for identifying spatially variable genes in spatial transcriptomics.
• Current approaches often compromise kernel expressiveness for computational tractability.

Purpose of the Study:

• To introduce Flash-S, a novel method for spatial gene expression analysis.
• To enhance accuracy, calibration, and scalability in identifying spatially variable genes.

Main Methods:

• Flash-S utilizes Random Fourier Features and sparse sketching to perform multi-scale kernel testing in the frequency domain.
• The method handles zero-inflated data without constructing distance matrices.
• A kurtosis-corrected null hypothesis ensures calibration.

Main Results:

• Flash-S achieved a mean Kendall's τ of 0.935 across 50 datasets from 9 platforms, outperforming existing methods.
• It processed the Allen Brain MERFISH atlas (3.94 million cells) in 12.6 minutes with 21.5 GB memory.
• Near-nominal false-positive rates were maintained under permutation testing.

Conclusions:

• Flash-S offers a scalable and accurate solution for spatial transcriptomics.
• The method successfully identified a mitochondrial biogenesis program in human cardiac tissue, which was missed by parametric methods.
• These findings were validated in an independent cohort, highlighting the method's robustness.