Accurately modeling biased random walks on weighted networks using node2vec.

Area of Science:

• Computational Biology
• Network Science
• Machine Learning

Background:

• Network embedding is crucial for machine learning on biological networks.
• Node2vec is a popular unsupervised method for network embedding using biased random walks.
• Existing methods like node2vec do not fully utilize edge weights in dense, weighted biological networks.

Purpose of the Study:

• To introduce node2vec+, an extension of node2vec that effectively incorporates edge weights into the random walk process.
• To evaluate the performance of node2vec+ against node2vec on synthetic and real-world biological datasets.
• To demonstrate the utility of node2vec+ in gene function and disease prediction tasks.

Main Methods:

• Developed node2vec+, a novel algorithm extending node2vec to account for edge weights in biased random walks.
• Utilized synthetic datasets to empirically assess the robustness of node2vec+ to noise compared to node2vec.
• Applied node2vec+ to genome-scale functional gene networks for gene function and disease prediction.

Main Results:

• node2vec+ shows improved robustness to additive noise in weighted graphs compared to node2vec.
• node2vec+ demonstrates superior performance over node2vec in gene function and disease prediction tasks on weighted gene networks.
• Both node2vec and node2vec+ outperform graph neural networks (GCN, GraphSAGE) in gene classification tasks with limited training data.

Conclusions:

• node2vec+ is a more effective network embedding method for weighted biological networks than node2vec.
• The incorporation of edge weights significantly enhances performance in biological network analysis, particularly for gene function and disease prediction.
• node2vec+ offers a valuable tool for analyzing complex biological networks, especially when data is scarce.