Leveraging heterogeneous network embedding for metabolic pathway prediction.

Area of Science:

• Systems Biology
• Bioinformatics
• Computational Biology

Background:

• Metabolic pathway reconstruction is crucial for understanding cellular function and potential.
• Current gene-centric methods often rely on extensive reference databases.
• Pathway-centric approaches offer hypothesis generation but require feature information.

Purpose of the Study:

• To introduce pathway2vec, a novel software package for automated feature generation in metabolic pathway inference.
• To develop a layered network architecture for capturing compound, enzyme, and pathway relationships.
• To create a low-dimensional embedding space for metabolic features using neural networks.

Main Methods:

• Developed pathway2vec, a software package with six representational learning modules.
• Constructed a three-layered network (compounds, enzymes, pathways) capturing inter- and betweenness interactions.
• Utilized neural embeddings to learn a low-dimensional representation of metabolic features.
• Benchmarked performance via node-clustering, embedding visualization, and pathway prediction against MetaCyc.

Main Results:

• The pathway2vec architecture effectively captures relevant biological relationships.
• Learned embeddings facilitate improved outcomes in metabolic pathway prediction tasks.
• Node-clustering and embedding visualization demonstrate the utility of the learned feature space.

Conclusions:

• Pathway2vec provides a powerful engine for hypothesis generation in systems biology.
• The generated embeddings can enhance the accuracy of metabolic pathway inference.
• This approach advances computational methods for analyzing cellular metabolic potential.