BioNEV

BioNEV evaluates and benchmarks graph embedding methods for analyzing biomedical networks to support link prediction and node classification tasks including drug-disease association (DDA), drug-drug interaction (DDI), protein-protein interaction (PPI), medical term semantic type classification, and protein function prediction.

Key Features:

• Systematic Evaluation: Systematically evaluates 11 representative graph embedding methods spanning matrix factorization, random walk-based, and neural network-based approaches.
• Matrix Factorization Methods: Includes specific matrix factorization techniques such as Laplacian Eigenmap, SVD (Singular Value Decomposition), Graph Factorization, HOPE, and GraRep.
• Benchmark Datasets: Provides five benchmark datasets tailored for biomedical prediction tasks including DDA, DDI, PPI, medical term semantic type classification, and protein function prediction.
• Comprehensive Comparisons: Compares embedding methods on three link prediction tasks (DDA, DDI, PPI) and two node classification tasks, assessing relative performance against traditional techniques.
• Performance Insights: Reports that recent graph embedding methods can achieve competitive results without relying on biological features.
• Guidelines and Hyper-parameter Tuning: Summarizes experimental outcomes to provide general guidelines for method selection and hyper-parameter settings for specific biomedical tasks.

Scientific Applications:

• Drug Discovery: Supports drug-disease association prediction to aid identification of potential therapeutic targets.
• Drug Interaction Analysis: Supports drug-drug interaction prediction to inform pharmacological studies.
• Protein Interaction Analysis: Supports protein-protein interaction prediction to improve understanding of cellular processes and disease mechanisms.
• Biomedical Network Classification: Supports medical term semantic type classification and protein function prediction to enhance biological annotation.

Methodology:

Implements matrix factorization-based methods (Laplacian Eigenmap, SVD (Singular Value Decomposition), Graph Factorization, HOPE, GraRep), random walk-based embeddings, and neural network-based embeddings within a structured benchmarking framework to evaluate performance on curated datasets and compare methods including assessments of operation without biological features.