MGDMCL: A multi-omics integration framework based on masked graph dynamic learning and multi-granularity feature contrastive learning for biomedical classification.

BACKGROUND AND OBJECTIVE

Integrating multi-omics data facilitates a comprehensive understanding of the etiology of complex diseases, which is critical for achieving precision medicine. Recently, graph-based approaches have been increasingly leveraged in the integrative multi-omics data analysis due to their robust expressive capability. However, these methods still face two limitations: 1) relying predominantly on a fixed sample similarity graph (SSG) to obtain omics-specific feature representation, and 2) insufficiently exploring the interrelations between different features from various omics. To this end, we propose MGDMCL, an innovative framework for integrating multiple omics data based on masked graph dynamic learning and multi-granularity feature contrastive learning.

METHODS

For each type of omics data, a masked graph dynamic learning approach adaptively adjusts the SSG structure and achieves the learning of a reliable SSG in a graph dynamic learning manner, obtaining multi-layer feature representations from various graph convolutional networks (GCN) layers. Then, the multi-layer feature representations of different omics are concatenated at the layer-level, and a multi-granularity feature contrastive learning is designed to learn consensus feature representations of specific layers. Furthermore, to enhance classification robustness, the true class probability is introduced to evaluate the classification confidence of consensus feature representations from different layers.

RESULTS

Extensive experiments on five public datasets, including LGG, ROSMAP, LUSC, BRCA, and KIPAN, show that MGDMCL significantly surpasses state-of-the-art baselines in various biomedical classification tasks.

CONCLUSIONS

The proposed MGDMCL provides a more effective approach for integrative multi-omics data analysis, exhibiting great potential in biomedical classification applications. The implementation code of MGDMCL has been released at https://www.github.com/wxchen-uestc/MGDMCL.