In Silico Optimization of Tissue Microarray Design for Machine Learning Analysis

Machine learning (ML) methods based on histopathology slides are increasingly used for cancer diagnosis and treatment. Analyses with specialized experimental stains, such as spatial proteomics, are often limited to tissue microarrays (TMAs) rather than whole slide images due to the cost of data generation. While different approaches for selecting tissue to include in a TMA exist, the impact of the sampling protocol on downstream ML-model performance is under-explored. Here, we present a general-purpose framework to evaluate ML performance of synthetic TMA datasets. We compare the performance of 16 synthetic TMA breast carcinoma datasets for predicting patient overall survival, immunohistochemistry subtype, and mutation status. We deduce several common design principles that are broadly applicable across TMA applications. Finally, we conduct a case study into principled and task-specific TMA sampling procedures in budget-constrained scenarios.