DEDUCE: Multi-head attention decoupled contrastive learning to discover cancer subtypes based on multi-omics data

Due to the high heterogeneity and clinical characteristics of cancer, there are significant differences in multi-omics data and clinical features among subtypes of different cancers. Therefore, the identification and discovery of cancer subtypes are crucial for the diagnosis, treatment, and prognosis of cancer. In this study, we proposed a generalization framework based on attention mechanisms for unsupervised contrastive learning to analyze cancer multi-omics data for the identification and characterization of cancer subtypes. The framework contains a symmetric unsupervised multi-head attention encoder, which can deeply extract contextual features and long-range dependencies of multi-omics data, reducing the impact of noise in multi-omics data. Importantly, the proposed framework includes a decoupled contrastive learning model (DEDUCE) based on a multi-head attention mechanism to learn multi-omics data features and clustering and identify cancer subtypes. This method clusters subtypes by calculating the similarity between samples in the feature space and sample space of multi-omics data. The basic idea is to decouple different attributes of multi-omics data features and learn them as contrasting terms. Construct a contrastive loss function to measure the difference between positive examples and negative examples, and minimize this difference, thereby encouraging the model to learn better feature representation. The DEDUCE model conducts large-scale experiments on simulated multi-omics data sets, single-cell multi-omics data sets and cancer multi-omics data sets, and the results are better than 10 deep learning models. Finally, we used the DEDUCE model to reveal six cancer subtypes of AML. By analyzing GO functional enrichment, subtype-specific biological functions and GSEA of AML,