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

Area of Science:

• Computational biology
• Bioinformatics
• Machine learning in oncology

Background:

• Cancer exhibits high heterogeneity and clinical diversity, leading to variations in multi-omics data and clinical features across subtypes.
• Accurate identification and characterization of cancer subtypes are crucial for personalized treatment strategies.

Purpose of the Study:

• To develop an unsupervised contrastive learning model, DEDUCE, for analyzing multi-omics cancer data.
• To identify and characterize distinct cancer subtypes by extracting deep contextual features and long-range dependencies.

Main Methods:

• DEDUCE employs symmetric multi-head attention encoders (SMAE) for unsupervised feature extraction from multi-omics data.
• A subtype-decoupled contrastive learning method with a multi-head attention mechanism is used for simultaneous feature learning and clustering.
• Clustering is performed by calculating sample similarity in both feature and sample spaces, optimizing a contrastive loss function.

Main Results:

• DEDUCE demonstrated superior performance against 10 deep learning models on simulated, single-cell, and cancer multi-omics datasets.
• Ablation experiments confirmed the effectiveness of individual modules within the DEDUCE model.
• The model successfully identified six distinct cancer subtypes in Acute Myeloid Leukemia (AML).

Conclusions:

• The DEDUCE model effectively learns features from multi-omics data using SMAE and subtype-decoupled contrastive learning for robust cancer subtype identification.
• DEDUCE shows significant potential in discovering novel cancer subtypes and enhancing interpretability through functional enrichment analysis.
• Application to AML identified six subtypes, with further analysis revealing subtype-specific biological functions and pathways.