Deep learning generates custom-made logistic regression models for explaining how breast cancer subtypes are...

Area of Science:

• Computational biology
• Genomics
• Medical informatics

Background:

• Accurate breast cancer intrinsic subtype differentiation is vital for treatment planning.
• Deep learning excels at subtype prediction but lacks interpretability regarding gene associations.

Purpose of the Study:

• To develop an explainable deep learning model (PWL) for breast cancer subtype analysis.
• To identify specific genes linked to intrinsic breast cancer subtypes and elucidate underlying mechanisms.

Main Methods:

• Developed a point-wise linear (PWL) model, an explainable deep learning approach generating patient-specific logistic regressions.
• Trained the PWL model using RNA-seq data for PAM50 intrinsic subtype prediction.
• Employed a deep enrichment analysis to explore relationships between subtypes and copy number variations.

Main Results:

• The PWL model successfully predicted PAM50 intrinsic subtypes using RNA-seq data.
• Analysis revealed that the PWL model identified genes associated with cell cycle-related pathways.
• The model demonstrated the clinical utility of analyzing breast cancer subtypes.

Conclusions:

• The PWL model offers an interpretable deep learning strategy for breast cancer subtype analysis.
• This approach can uncover mechanisms driving breast cancer subtypes and potentially improve clinical outcomes.
• The study highlights the potential of explainable AI in precision oncology.