CautionGCN: Cancer Subtype Classification by Developing Causal Multi-Head Autoencoder and Graph Convolutional Network
Na Li, Fan Yang, Qingke Zhang, Q. X. Li, Xiaobin Zhang, Jiayi Teng · 2024
Cancer is an aggressive and complex disease, and its heterogeneity makes it challenging to target specific therapies for different tumor types. Predicting cancer subtypes using only single-omics data is challenging because cancer occurrence and development result from the concerted action of multiple molecular mechanisms, not a singular factor. Recently, Deep Learning approaches have become a valuable tool in cancer subtype classification. However, high-dimensional multi-omics data are typically imbalanced, with an abundance of molecular features and relatively few patient samples. This imbalance makes it difficult to conduct integrated multi-omics analysis based on deep learning. In this study, CautionGCN, a model using deep neural networks based on multi-omics data, was employed to classify breast cancer subtypes. Three types of breast cancer omics data-RNAseq, DNA methylation, and copy number variation-were collected from The Cancer Genome Atlas. The proposed model comprises the following three components. First, a similarity network among the three omics samples was constructed using a similarity network fusion method. The causal Multi-Head autoencoder then extracts the causal features of the three omics data. Finally, the sample similarity network and feature matrix were input into a graph convolutional neural network for cancer subtype classification. The robustness and classification accuracy of the model can be enhanced by identifying genetic features with causal relationships. Experimental results confirmed that our model has significantly improved performance compared to other methods, and the effectiveness of multi-omics integration and causal feature selection was identified.