Histopathology-based protein multiplex generation using deep learning

Area of Science:

• Computational biology
• Pathology
• Artificial intelligence in medicine

Background:

• Multiplexed protein imaging is crucial for understanding tumor-microenvironment interactions but faces limitations in cost, time, and tissue accessibility.
• Standard hematoxylin and eosin (H&E) histopathology images are widely available but lack detailed protein information.

Purpose of the Study:

• To develop a deep learning framework, HistoPlexer, for generating spatially resolved protein multiplexes from standard H&E images.
• To enable cost- and time-efficient characterization of the tumor microenvironment for advancing precision oncology.

Main Methods:

• HistoPlexer utilizes a conditional generative adversarial network architecture with custom loss functions.
• The framework jointly predicts multiple tumor and immune markers, ensuring pixel- and embedding-level similarity while minimizing slice-to-slice variations.
• Validation involved expert assessment on metastatic melanoma samples and benchmarking on diverse, publicly available cancer datasets.

Main Results:

• HistoPlexer-generated protein maps closely resemble experimentally derived maps and preserve key biological relationships, including protein co-localization patterns.
• The predicted immune infiltration patterns enabled accurate stratification of tumors into distinct immune subtypes.
• Integration of HistoPlexer-derived features improved survival prediction and immune subtype classification models compared to using H&E features alone.
• The method demonstrated robustness and outperformed baseline approaches across various cancer types and imaging conditions.

Conclusions:

• HistoPlexer provides a powerful, efficient method for whole-slide protein multiplex generation from routine H&E images.
• This approach significantly enhances tumor microenvironment characterization, offering a valuable tool for precision oncology.
• The framework's ability to stratify tumors and improve predictive model performance highlights its potential to impact clinical decision-making.