Predicting DNA replication-related protein expression at the nucleus level from HE: A deep learning study with...

Area of Science:

• Digital Pathology
• Computational Pathology
• Biomedical Imaging

Background:

• Hematoxylin-eosin (HE) staining is a fundamental technique in histopathology.
• Ancillary testing, often involving immunohistochemistry (IHC), is crucial for molecular subtyping and treatment decisions.
• Reducing the burden of ancillary testing can streamline diagnostic workflows.

Purpose of the Study:

• To evaluate the capability of deep convolutional networks to predict per-nucleus protein expression directly from HE images.
• To assess the prediction of continuous optical density (OD) and binary positivity for DNA replication-related proteins (CDC6, CDT1, MCM7, ORC1, CDC7, Geminin) and Ki-67.
• To determine if molecular information can be inferred from HE nuclear morphology, thereby reducing the need for IHC.

Main Methods:

• A paired HE/IHC dataset was constructed from 21 endometrioid carcinoma cases.
• Nuclear segmentation was performed using HoVer-Net, and HE images were processed for color unmixing.
• Deep learning models (ResNet-50, EfficientNet-B0, MobileNetV3-Small) were trained to predict per-nucleus OD and positivity from HE nuclear crops.

Main Results:

• Per-nucleus protein expression prediction from HE morphology was feasible with moderate performance.
• MCM7 and Ki-67 showed the strongest discrimination (AUC-ROC ≈ 0.70-0.72).
• The Ki-67 labeling index showed moderate agreement with whole-slide imaging (WSI)-based digital IHC (r ≈ 0.54).

Conclusions:

• Protein expression can be inferred from HE nuclear morphology alone, suggesting potential clinical utility.
• These findings support the development of AI-driven tools for digital pathology.
• Future research should focus on larger cohorts and external validation to confirm these promising results.