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

Optimizing genomic-based molecular subtyping is key to promote personalized medicine. While neural networks face setbacks regarding tabular data modeling, deep learning has undergone groundbreaking advances across multiple domains, catalyzing further breakthroughs across AI applications. New neural network architectures exhibit enhanced performance, efficiency, and robustness, which could benefit the genomic use-case. In this study we introduce SubNExT, an optimized shallow CNN with a ConvNeXt backbone using t-SNE and DeepInsight 2D-converted gene expression for breast cancer subtyping. It was compared with other modelization strategies for gene expression data, by optimizing a Transformer, an MLP and XGBoost for unconverted values, a 1D CNN (NeXt-TDNN) for ordered values, and a ViT as an alternative for 2D-converted expression. During evaluation, SubNExT obtains an accuracy of 87.12%, matching the state-of-the-art XGBoost and its 87.24% acc at the top of the benchmark. SubNExT manages this performance with just 76k parameters and the shortest training time, as well as the best stability and robustness among all considered approaches. By providing accurate, efficient and robust molecular subtyping of breast cancer using gene expression data, SubNExT and its design principles catalyze deep learning adoption in oncogenomics.