Benchmarking sketching methods on spatial transcriptomics data

Area of Science:

• Computational biology
• Genomics
• Bioinformatics

Background:

• High-throughput spatial transcriptomics (ST) generates massive datasets, posing computational challenges for analysis.
• Current sub-sampling methods (sketching) often prioritize gene expression, neglecting physical location and introducing spatial bias.
• Existing methods risk distorting tissue architecture by over-sampling highly variable regions and under-sampling homogeneous areas.

Purpose of the Study:

• To systematically benchmark existing sketching methods for spatial transcriptomics.
• To evaluate the impact of different data representations (PCA embeddings, spatial coordinates, smoothed embeddings) on sampling accuracy.
• To develop and validate a novel sketching approach that balances transcriptomic representation with spatial integrity.

Main Methods:

• Benchmarking of uniform sampling, leverage-score sampling, Geosketch, and scSampler across diverse ST datasets (mouse ovary, MERFISH brain, human breast cancer, lung) and simulations.
• Input representations included PCA embeddings, spatial coordinates, and spatially smoothed embeddings.
• Development of a spatially aware method using leverage scores from a spatially smoothed randomized SVD basis.

Main Results:

• Expression-only sketching captures global heterogeneity but distorts tissue architecture.
• Coordinate-only sampling preserves tissue coverage but misses transcriptional extremes.
• Spatially smoothed leverage scores demonstrated superior performance in maintaining tissue coverage, recovering rare cell states, and avoiding edge effects, outperforming alternatives across multiple metrics (Hausdorff distance, ARI, PCA loading drift, MSE).

Conclusions:

• Standard sketching methods for ST are prone to spatial bias.
• A novel, spatially aware sketching approach using smoothed leverage scores effectively balances transcriptomic and spatial information.
• This method enables fast, unbiased analysis of large-scale spatial transcriptomics data, preserving both cellular heterogeneity and tissue architecture.