Clustering approaches for visual knowledge exploration in molecular interaction networks.

Area of Science:

• Biomedical informatics
• Computational biology
• Data mining

Background:

• Biomedical knowledge is increasingly complex and stored in network repositories.
• Analyzing large, structured biomedical graphs visually presents computational challenges.

Purpose of the Study:

• To investigate knowledge discovery in molecular interaction diagrams using various distance metrics for clustering.
• To evaluate the impact of combining distance metrics and their order on clustering accuracy and biological knowledge discovery.

Main Methods:

• Utilized Euclidean, shortest path, and ontology-based distances for clustering molecular interaction diagrams.
• Employed hierarchical and a novel bi-level optimization approach with evolutionary algorithms for distance metric combination.
• Assessed cluster quality by comparing with expert knowledge and analyzing enrichment of Gene and Disease Ontology terms.

Main Results:

• Combining distance metrics improved clustering accuracy compared to expert-provided clusters.
• The performance of distance metric combinations was dependent on clustering depth (number of clusters).
• Bi-level optimization identified the relative importance of distance functions and their order in clustering.
• Both hierarchical and bi-level clustering discovered more Gene and Disease Ontology terms than expert clusters.
• Bi-level clustering outperformed hierarchical clustering in several instances.

Conclusions:

• Combining distance functions enhances clustering and exploration of visual biomedical knowledge repositories.
• Bi-level optimization effectively evaluates the importance of distance function order in clustering.
• Both the combination and order of distance functions significantly impact clustering quality and knowledge recognition.
• Simultaneous utilization of multiple dimensions (distance metrics) is proposed for visual knowledge exploration.