ROSIE: AI generation of multiplex immunofluorescence staining from histopathology images.

Area of Science:

• Computational pathology
• Biomarker discovery
• Artificial intelligence in medicine

Background:

• Hematoxylin and eosin (H&E) staining is a fundamental histopathology technique.
• H&E lacks direct molecular marker information, necessitating further assays.
• Identifying specific cell types and microenvironments is crucial for disease understanding.

Purpose of the Study:

• To introduce ROSIE, a deep learning framework for imputing protein expression and localization from H&E images.
• To enable in silico multiplex immunofluorescence (mIF) analysis.
• To enhance the diagnostic and prognostic capabilities of routine histopathology.

Main Methods:

• Training a deep learning model on over 1300 paired H&E and mIF samples across diverse tissues and diseases.
• Utilizing a dataset spanning over 16 million cells for robust model development.
• Validating the imputation accuracy on held-out H&E samples.

Main Results:

• ROSIE accurately predicts the expression and localization of numerous proteins from H&E images.
• In silico mIF effectively distinguishes cell phenotypes, including lymphocytes (B cells, T cells).
• The framework robustly identifies stromal, epithelial microenvironments, and tumor-infiltrating lymphocytes (TILs).

Conclusions:

• ROSIE bridges the gap between H&E morphology and molecular information.
• The computational approach aids in understanding tumor-immune interactions and cancer progression.
• This AI-driven tool can inform novel treatment strategies and improve cancer diagnostics.