Statistical analysis of multiple regions-of-interest in multiplexed spatial proteomics data

Area of Science:

• Computational pathology
• Cancer research
• Biostatistics

Background:

• Multiplexed spatial proteomics reveals tumor cell organization, impacting survival and treatment response.
• Spatial summary statistics like Ripley's K and Besag's L quantify this organization.
• Synthesizing data from multiple images per patient for clinical endpoint association remains challenging.

Purpose of the Study:

• To evaluate existing and novel methods for associating spatial summary statistics from multiple tumor images with patient-level clinical outcomes.
• To determine the most effective strategy for integrating multi-image spatial proteomics data.

Main Methods:

• Evaluation of averaging-based approaches (weighted mean of summary statistics).
• Proposal and evaluation of ensemble testing approaches using simulated random weights and P-value aggregation.
• Systematic performance evaluation via simulation and application to non-small cell lung cancer, colorectal cancer, and triple-negative breast cancer data.

Main Results:

• A simple weighted average of summary statistics, weighted by cell count per image, demonstrated high power and effective type I error control.
• Incorporating image region size variation into weighted aggregation can improve power when size is informative.
• Ensemble testing showed high power and type I error control across various simulated conditions.

Conclusions:

• The optimal strategy for multi-image spatial proteomics data integration depends on the specific dataset.
• Weighted averaging based on cell counts is a robust and often superior method.
• Ensemble testing offers a powerful alternative, particularly when image characteristics vary.