Breast Cancer Subtype Classification via Multi-Omics Integration Using Bayesian Optimization and Contrastive Learning

Cancer remains a leading cause of death in Japan, highlighting the urgent need for further advancements in cancer treatment. Breast cancer, in particular, can be classified into several distinct subtypes based on unique gene expression patterns, and such subtype information is crucial for personalized cancer therapy. To improve the accuracy of subtype prediction, there has been an increasing emphasis on multi-omics data analysis. However, integrating multiple types of omics data effectively continues to pose a significant challenge. In this study, we propose a novel integration method utilizing deep learning to individually extract features from three types of omics data—RNA, miRNA, and DNA methylation (Meth)—obtained from publicly available databases. These extracted features are subsequently integrated through matrix multiplication, where Bayesian optimization dynamically adjusts the dimensionality of features based on the relative importance of each omics type. Additionally, we incorporate contrastive learning into the encoder structure to further improve classification accuracy. Our proposed method achieved an accuracy of 0.822 and a macro-average F1 score of 0.798, indicating superior performance and practicality compared with integration models reported in previous studies.