Graph2MDA

Graph2MDA predicts microbe–drug associations using a multi-modal variational graph embedding model to identify potential microbial influences on drug efficacy and toxicity.

Key Features:

• Variational Graph Autoencoder (VGAE): Processes multi-modal attributed graphs to embed nodes and graphs using a variational graph autoencoder architecture.
• Multi-Modal Attributed Graphs: Integrates molecular structures, genetic sequences, and functional annotations of microbes and drugs into attributed graph representations.
• Latent Representation Learning: Learns informative latent representations for individual nodes (microbes and drugs) and whole graphs to capture underlying interaction patterns.
• Deep Neural Network Classifier: Predicts potential microbe–drug associations using a deep neural network trained on the learned latent features.
• Robustness and Sensitivity Analysis: Includes hyperparameter analysis and model ablation studies to assess model sensitivity and robustness.
• Performance Evaluation: Evaluated on three independent datasets, compared against six state-of-the-art methods, with drug clustering consistent with the Anatomical Therapeutic Chemical (ATC) classification and case studies showing 75–95% concordance with PubMed reports.

Scientific Applications:

• Drug Research and Development: Identifies candidate microbe–drug associations to support discovery of antibacterial agents and to inform drug efficacy and toxicity assessments.
• Microbial Genomics and Pharmacology: Leverages large-scale microbial genomic and pharmacological datasets to uncover potential therapeutic targets and microbe-mediated drug interactions.

Methodology:

Constructs multi-modal attributed graphs from molecular structures, genetic sequences, and functional annotations of microbes and drugs; applies a variational graph autoencoder (VGAE) to learn latent node and graph representations; and uses a deep neural network classifier to predict associations, with hyperparameter analysis, model ablation studies, and evaluation on three independent datasets including ATC-based clustering and PubMed case-study comparison.