Completing spatial transcriptomics data for gene expression prediction benchmarking

Area of Science:

• Computational Biology
• Genomics
• Biotechnology

Background:

• Spatial transcriptomics integrates histology with gene expression but faces challenges like high cost and data dropout.
• Deep learning models predict gene expression from histology images, but lack standardized evaluation due to dataset and protocol inconsistencies.

Purpose of the Study:

• To establish a standardized resource for evaluating gene expression prediction models.
• To develop an advanced model for completing gene expression data.
• To create a comprehensive benchmark for spatial transcriptomics research.

Main Methods:

• Curated 26 public datasets into the SpaRED database for standardized model evaluation.
• Developed SpaCKLE, a transformer-based model for gene expression completion.
• Established the SpaRED benchmark, evaluating eight prediction models on raw and completed data.

Main Results:

• SpaRED provides a standardized resource for spatial transcriptomics model evaluation.
• SpaCKLE reduced mean squared error by over 82.5% in gene expression completion.
• SpaCKLE significantly improved performance across all evaluated gene expression prediction models.

Conclusions:

• The SpaRED database and SpaCKLE model offer a comprehensive benchmark and improved methodology for gene expression prediction from histology images.
• These contributions facilitate more reliable and accessible spatial transcriptomics research, paving the way for clinical applications.