Patient Graph Deep Learning to Predict Breast Cancer Molecular Subtype.

Area of Science:

• Oncology
• Bioinformatics
• Machine Learning

Background:

• Breast cancer is a complex disease with diverse genomic mutations and clinical features.
• Molecular subtypes significantly influence patient prognosis and treatment decisions.
• Accurate subtype prediction is crucial for effective breast cancer management.

Purpose of the Study:

• To develop a deep graph learning framework for integrating multimodal patient data.
• To enhance the representation of breast cancer patient information.
• To accurately predict molecular subtypes of breast cancer.

Main Methods:

• Constructed a multi-relational directed graph representing patient data from multiple disciplines.
• Extracted radiographic features from DCE-MRI using a dedicated pipeline.
• Employed an autoencoder for genomic variant embedding.
• Utilized transfer learning and a Relational Graph Convolutional Network (RGCN) for subtype prediction.

Main Results:

• Multimodal data integration significantly improved breast cancer molecular subtype prediction.
• The model generated more distinct and informative feature representations for patients.
• Demonstrated the effectiveness of deep learning for multimodal data fusion in oncology.

Conclusions:

• Deep graph learning offers a powerful approach for fusing diverse breast cancer data.
• This method enhances the understanding and prediction of molecular subtypes.
• The findings support the potential of AI in advancing personalized breast cancer care.