Bind: large-scale biological interaction network discovery through knowledge graph-driven machine learning.

Area of Science:

• Computational Biology
• Bioinformatics
• Network Science

Background:

• Biological systems comprise complex, interconnected networks crucial for disease understanding and therapeutics.
• Current AI interaction predictors are isolated, missing holistic network effects.
• Wet-lab validation is costly and time-intensive, necessitating advanced computational tools.

Purpose of the Study:

• To develop a unified platform for predicting and analyzing diverse biological interactions.
• To overcome limitations of isolated prediction methods and wet-lab approaches.
• To facilitate comprehensive biological network analysis for therapeutic development.

Main Methods:

• Developed BIND (Biological Interaction Network Discovery), a framework using 11 Knowledge Graph Embedding methods.
• Employed a two-stage training strategy to address class imbalance and heterogeneity.
• Integrated entity embeddings into 7 machine learning classifiers, creating 1,050 predictive pipelines.

Main Results:

• Simpler embedding models effectively captured biological patterns, often surpassing complex methods.
• The two-stage training improved protein-protein interaction prediction by up to 26.9%.
• Optimal BIND pipelines achieved high F1-scores (0.85-0.99); 1355 high-confidence predictions were generated in a drug-phenotype case study.

Conclusions:

• BIND offers a unified web application for simultaneous prediction and analysis of multiple biological interaction types.
• The platform demonstrates superior performance over isolated methods for biological network analysis.
• BIND accelerates biomarker discovery and therapeutic development by enabling experimental validation of novel interactions.