BCtypeFinder: A Semi-Supervised Model with Domain Adaptation for Breast Cancer Subtyping Using DNA Methylation...

Area of Science:

• Genomics
• Epigenetics
• Computational Biology

Background:

• Accurate breast cancer subtype prediction is crucial for personalized medicine.
• DNA methylation patterns are emerging as key epigenetic markers for distinct breast cancer subtypes.
• Challenges exist in developing reliable prediction models due to limited annotated datasets and batch effects.

Purpose of the Study:

• To develop an advanced computational framework, BCtypeFinder, for accurate breast cancer subtype prediction.
• To address the scarcity of labeled data and batch effects in DNA methylation profiling.
• To leverage both labeled and unlabeled data for robust subtype classification.

Main Methods:

• Utilized a domain adaptation network combined with semi-supervised learning.
• Employed DNA methylation profiles for feature extraction.
• Implemented batch correction techniques to mitigate dataset-specific variations.

Main Results:

• BCtypeFinder demonstrated superior classification performance compared to existing methods across multiple test cases.
• The framework successfully extracted domain-invariant features and aligned subtype distributions.
• Batch correction within BCtypeFinder effectively removed patient batch variations within subtypes, enhancing classifier robustness.

Conclusions:

• BCtypeFinder offers a robust and accurate solution for breast cancer subtype prediction using DNA methylation data.
• The integration of domain adaptation and semi-supervised learning effectively handles data heterogeneity.
• Public availability of BCtypeFinder facilitates further research and clinical application.