SOFisher: reinforcement learning-guided experiment designs for spatial omics

Area of Science:

• Spatial omics
• Computational biology
• Biotechnology

Background:

• Spatial omics technologies precisely map molecules like proteins and RNAs within tissues.
• Current field of view (FOV) sampling in spatial omics is often inefficient, requiring dense acquisition and stitching.
• Optimizing FOV selection is critical for maximizing data yield and minimizing resource expenditure.

Purpose of the Study:

• To develop an intelligent framework, SOFisher, for optimizing FOV sampling strategies in spatial omics experiments.
• To enhance the efficiency of capturing regions of interest using reinforcement learning.
• To enable deeper biological insights from reduced experimental footprints.

Main Methods:

• SOFisher, a reinforcement learning-based framework, guides sequential FOV selection.
• The framework learns from previously sampled FOVs to predict optimal next positions.
• Performance was evaluated using simulations on real spatial datasets and cross-domain generalization tests.

Main Results:

• SOFisher consistently outperformed conventional sampling strategies across various metrics.
• The framework demonstrated robustness and generalizability across different FOV sizes and datasets.
• Application on Alzheimer's Disease and colorectal cancer datasets revealed key biological insights, including cell states and gene programs, from limited FOVs.

Conclusions:

• SOFisher significantly improves the efficiency and effectiveness of spatial omics experimental design.
• The AI-driven approach can yield comprehensive biological insights comparable to traditional methods using fewer resources.
• SOFisher has the potential to revolutionize spatial biology experiment design, making complex analyses more accessible.