Weakly-supervised tumor purity prediction from frozen H&E stained slides

Area of Science:

• Computational pathology
• Digital pathology
• Cancer genomics

Background:

• Accurate tumor purity estimation is crucial for precision medicine and reliable next-generation sequencing analysis.
• Current molecular-based purity estimation methods are costly and time-consuming due to required tumor sequencing.
• Hematoxylin and eosin (H&E) stained histological slides offer a readily available data source for purity estimation.

Purpose of the Study:

• To develop and validate a novel computational approach for accurate tumor purity quantification directly from digital H&E stained slides.
• To establish a weakly-supervised method that reduces the reliance on extensive manual annotation or fully-supervised learning.

Main Methods:

• Development of a weakly-supervised purity (wsPurity) model utilizing deep learning techniques.
• Training and validation of the wsPurity model on diverse cancer types from The Cancer Genome Atlas (TCGA).
• Evaluation of model performance on independent cohorts, including external datasets and frozen slides.

Main Results:

• The wsPurity model demonstrated high accuracy in predicting tumor purity on unseen TCGA data.
• The model exhibited strong generalizability, achieving an F1-score of 0.83 for prostate adenocarcinoma in an external cohort.
• wsPurity outperformed a comparable fully-supervised approach and showed improved generalizability to unseen frozen slides (0.1543 Mean Absolute Error).
• The model successfully identified high-resolution tumor regions and enabled stratification of tumors based on purity levels.

Conclusions:

• The developed deep learning model effectively analyzes tumor H&E sections for purity estimation.
• The wsPurity model is generalizable to diverse H&E stained slides from different data sources, including TCGA and Weill Cornell Medicine.
• This approach provides a valuable, efficient tool for tumor purity assessment in clinical and research settings.