MOCapsNet: Multiomics Data Integration for Cancer Subtype Analysis Based on Dynamic Self-Attention Learning and Capsule Networks

Background and Objective: With the rapid development of the accumulation of large-scale multiomics data sets, integrating various omics data to provide a thorough study from multiple perspectives can significantly provide stronger support for precise treatment of diseases. However, due to the complexity of multiomics data, issues of feature redundancy and noise often do not receive sufficient attention when processing high-dimensional data. Moreover, simple concatenation strategies may overlook the correlations between different omics data, failing to effectively capture the unique information inherent in multiomics data. Meanwhile, deep neural networks often rely on complex structures and numerous parameters for training and inference, making their internal feature representations and decision-making processes difficult to interpret. Methods: We propose an interpretable multiomics data integration method for cancer subtype classification, named MOCapsNet, based on self-attention and capsule networks. Specifically, the self-attention confidence learning module is implemented to assess the feature information within each omic and to assign weights to the embedded representations of various groups, resulting in more targeted integrated information. Furthermore, the capsule network structure is employed for the final cancer classification task. Results: The model achieved strong performance on both tasks: 87.8% accuracy on the BRCA multiclassification data set and 83.6% accuracy with an AUC of 88.8% on the LGG data set. Conclusions: The proposed framework has undergone extensive testing on omics data sets, consistently proving its effectiveness in integrating multiomics data. It improves classification accuracy while enhancing the interpretability of results by fully utilizing the feature information.