SubNExT: Towards accurate, efficient and robust gene expression classification for breast cancer subtyping

Area of Science:

• Genomics
• Artificial Intelligence
• Computational Biology

Background:

• Optimizing genomic-based molecular subtyping is crucial for advancing personalized medicine.
• Deep learning (DL) models have shown significant advancements, offering potential for complex biological data analysis.
• Traditional neural networks often struggle with tabular data, necessitating novel architectures for genomic applications.

Purpose of the Study:

• To introduce SubNExT, a novel deep learning model for breast cancer molecular subtyping.
• To evaluate SubNExT's performance against various established computational strategies using gene expression data.
• To demonstrate the efficacy of optimized shallow convolutional neural networks (CNNs) with advanced backbones in oncogenomics.

Main Methods:

• SubNExT utilizes a shallow CNN with a ConvNeXt backbone, processing t-SNE and DeepInsight 2D-converted gene expression data.
• Comparative analysis included optimized Transformer, MLP, XGBoost (unconverted values), NeXt-TDNN (1D CNN, ordered values), and Vision Transformer (ViT, 2D-converted expression).
• Performance was benchmarked using accuracy, parameter count, training time, stability, and robustness metrics.

Main Results:

• SubNExT achieved a high accuracy of 87.12%, comparable to the state-of-the-art XGBoost (87.24%).
• The model demonstrated superior efficiency with only 76,000 parameters and the shortest training duration.
• SubNExT exhibited the best stability and robustness among all evaluated approaches.

Conclusions:

• SubNExT provides an accurate, efficient, and robust method for molecular subtyping of breast cancer from gene expression data.
• The design principles of SubNExT encourage the adoption of deep learning techniques in oncogenomics.
• This study highlights the potential of advanced DL architectures for complex genomic data analysis in precision oncology.