Spatial information matters: are traditional imputation methods effective for spatial transcriptomics data?

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Spatially resolved transcriptomics (SRT) offers high-resolution spatial context for biological discovery.
• SRT datasets are often sparse with dropout events, hindering accurate interpretation.
• Existing imputation methods lack systematic benchmarking on new SRT technologies.

Purpose of the Study:

• To evaluate state-of-the-art (SOTA) imputation methods on emerging SRT platforms.
• To introduce a novel imputation method, SpaMean-Impute, for SRT data.
• To assess SpaMean-Impute's performance and computational efficiency.

Main Methods:

• Evaluated seven SOTA imputation methods across five SRT platforms and 23 datasets.
• Developed SpaMean-Impute, incorporating spatial information for dropout mitigation and detection.
• Benchmarked SpaMean-Impute against SOTA methods using metrics like ARI, NMI, AMI, and HOMO.

Main Results:

• No single SOTA method consistently excelled; most struggled with valid dropout identification.
• SpaMean-Impute significantly outperformed SOTA methods in imputation accuracy (e.g., 16.15% ARI improvement).
• SpaMean-Impute demonstrated superior computational efficiency, being ~33x faster and requiring ~1500 MB less memory than deep learning methods.

Conclusions:

• SpaMean-Impute is a highly effective and efficient method for imputing sparse SRT data.
• The method's ability to leverage spatial information addresses limitations of existing techniques.
• SpaMean-Impute offers a valuable tool for analyzing emerging high-resolution SRT datasets.