MoCaGCN: Cancer Subtype Classification by Developing Causal Graph Structure Learning

The classification of cancer subtypes is a crucial step towards more personalized treatment and has important implications for understanding the biological mechanisms of cancer subtypes. Multi-omits data present challenges in cancer research due to their high-dimensional feature spaces, significant noise, and limited sample sizes. However, the scale and heterogeneity of this data makes integration and classification a non-trivial task. Models based on correlations often contain confounding and redundant features, resulting in uninterpretability and weak generalization. Therefore, we proposed a novel multi-omits classification framework named Muti-omits Causal Graph Convolutional Network incorporated with causal structure learning and graph neural networks. Furthermore, the model can improve the performance and generalization capabilities of cancer classification by considering the prior knowledge of gene-gene interactions. This study constructed a more efficient framework for cancer subtype classification based on multi-omics data, which mainly include multi-omits data preprocessing, omits causal graph network construction, and multi-omits data integration methods. Finally, we demonstrated that our model outperforms the slate-of-the-art methods with fewer variables for higher accuracy on the real lung adenocarcinoma dataset from the cancer genome atlas program.