Obtaining spatially resolved tumor purity maps using deep multiple instance learning in a pan-cancer study.

Area of Science:

• Computational pathology
• Genomics
• Cancer research

Background:

• Tumor purity, the proportion of cancer cells in tissue, is crucial for genomic analysis.
• Manual estimation of tumor purity from H&E slides by pathologists is time-consuming, subjective, and often inaccurate.
• Existing methods lack correlation with genomic tumor purity, impacting downstream analysis.

Purpose of the Study:

• To develop and validate a deep multiple instance learning model for predicting tumor purity from digital histopathology slides.
• To assess the model's performance across diverse cancer cohorts.
• To provide a tool for objective and efficient tumor purity assessment.

Main Methods:

• A deep multiple instance learning model was trained on H&E-stained digital histopathology slides.
• The model predicted tumor purity and generated purity maps.
• Model performance was evaluated against genomic tumor purity values in eight TCGA cohorts and a local cohort.

Main Results:

• The deep learning model accurately predicted tumor purity, showing high consistency with genomic values.
• The model demonstrated robust performance across multiple cancer types and cohorts.
• Tumor purity maps revealed spatial variations within tissue sections, offering insights into the tumor microenvironment.

Conclusions:

• The developed deep learning model offers an objective and efficient method for predicting tumor purity from histopathology slides.
• This approach can significantly reduce pathologist workload and inter-observer variability in sample selection for genomic studies.
• The model's ability to provide spatial tumor purity information enhances understanding of the tumor microenvironment and facilitates more accurate genomic analyses.