DeepGene-BC: Deep Learning-Based Breast Cancer Subtype Prediction via Somatic Point Mutation Profiles

Area of Science:

• Genomics
• Computational Biology
• Oncology

Background:

• Traditional breast cancer molecular subtyping relies on transcriptomic profiles, facing challenges in robustness and clinical use.
• Somatic point mutations offer a stable genomic alternative but present challenges like high dimensionality and sparsity.
• Existing methods struggle to leverage the full predictive potential of sparse genomic mutation data.

Purpose of the Study:

• To develop a deep learning framework, deepGene-BC, for accurate breast cancer molecular subtyping using genomic mutation data.
• To overcome the limitations of high dimensionality and sparsity in mutation data for predictive modeling.
• To integrate pathway information and advanced deep learning techniques for improved subtyping.

Main Methods:

• Developed deepGene-BC, a deep learning framework combining pathway-informed feature selection and a hybrid neural network.
• Employed mutation recurrence filtering, pathway priors, and mutual information for feature refinement.
• Utilized a specialized hybrid neural network architecture to model linear, interactive, and nonlinear patterns in sparse data.

Main Results:

• DeepGene-BC achieved 77.3% overall accuracy and 75.2% average sensitivity on an independent TCGA breast cancer cohort.
• Demonstrated strong discriminative performance with a macro-averaged AU-ROC of 0.94 (95% CI: 0.92-0.96).
• The framework effectively distilled genome-wide mutations into a compact, interpretable feature set.

Conclusions:

• DeepGene-BC successfully integrates biologically informed feature engineering with deep learning for breast cancer stratification.
• The framework shows significant promise for non-invasive molecular subtyping and advancing precision oncology.
• This approach offers a robust and potentially more clinically applicable method for breast cancer subtyping.