CLCLSA: Cross-omics linked embedding with contrastive learning and self attention for integration with incomplete...

Area of Science:

• Genomics
• Bioinformatics
• Computational Biology

Background:

• Understanding complex genetic diseases requires integrating multi-omics data.
• Heterogeneous omics data provides limited views; simultaneous integration offers comprehensive insights.
• Unpaired multi-omics data due to cost and technical limitations poses a significant challenge.

Purpose of the Study:

• To develop a deep learning method for multi-omics data integration, specifically addressing incomplete datasets.
• To enhance the understanding of complex genetic diseases and phenotypes through robust data integration.
• To improve multi-omics data classification accuracy with missing data points.

Main Methods:

• Proposed Cross-omics Linked unified embedding with Contrastive Learning and Self Attention (CLCLSA) model.
• Utilized cross-omics autoencoders for feature representation learning.
• Employed multi-omics contrastive learning to maximize mutual information between omics layers.
• Incorporated feature-level and omics-level self-attention mechanisms for dynamic feature identification.
• Applied a Softmax classifier for multi-omics data classification.

Main Results:

• CLCLSA demonstrated promising performance in multi-omics data classification.
• The model effectively handled both complete and incomplete multi-omics datasets.
• Experimental results on four public datasets validated the proposed method's efficacy.

Conclusions:

• CLCLSA offers a robust solution for multi-omics integration with incomplete data.
• The method advances the potential for comprehensive disease etiology understanding.
• This approach facilitates more accurate classification of complex genetic diseases.