Multi-view clustering for multi-omics data using unified embedding.

Area of Science:

• Computational Biology
• Data Science
• Machine Learning

Background:

• Real-world datasets often feature multiple views, offering complementary information.
• Embedding learning is crucial for dimensionality reduction and nearest neighbor search in large datasets.

Purpose of the Study:

• To develop a unified embedding for multi-view data by learning a shared probability distribution.
• To optimally preserve neighborhood identity across different data views.
• To enable effective clustering of both complete and incomplete multi-view datasets.

Main Methods:

• A novel Multi-view Neighbourhood Embedding (MvNE) methodology is proposed.
• Probability distributions from each view are combined using conflation.
• Kullback-Leibler divergence serves as the cost function for gradient adjustment.
• The AMOSA multi-objective clustering technique is applied to the embedded space.

Main Results:

• The MvNE methodology successfully generates unified embeddings from multi-view data.
• It handles both complete and incomplete multi-view datasets effectively.
• Performance evaluation on 10 omics datasets shows a 2-3% improvement over state-of-the-art methods.

Conclusions:

• MvNE provides an effective approach for multi-view embedding and clustering.
• The method demonstrates superior performance compared to existing techniques in omics data analysis.