Multiplex Embedding of Biological Networks Using Cross-Network Node Similarities.

Area of Science:

• Computational Biology
• Machine Learning
• Network Science

Background:

• Network embedding is crucial for computational biology tasks.
• Existing methods often process multiplex networks (multiple related networks) independently.
• This overlooks valuable information from network topology and sparsity.

Purpose of the Study:

• To develop a novel multiplex network embedding technique that integrates multiple networks effectively.
• To account for inner-network smoothness and topological similarity between network layers.
• To improve node representations for downstream machine learning applications.

Main Methods:

• Formulated an optimization problem considering inner-network smoothness and topological similarity.
• Quantified topological similarity using shared neighborhood across networks.
• Developed an efficient iterative algorithm for computing diffusion states, improving runtime.
• Integrated diffusion states and used dimensionality reduction (SVD) for node embeddings.

Main Results:

• The proposed algorithm, Crossim, was evaluated for protein function prediction.
• Crossim embeddings consistently improved predictive accuracy compared to existing methods.
• Accounting for topological similarity across network layers enhances network integration.

Conclusions:

• Multiplex network embedding can be significantly improved by considering topological similarity.
• Crossim offers a more effective approach to integrating multiple biological networks.
• This method has broad implications for machine learning in computational biology.