NinimHMDA

NinimHMDA predicts multiple types of associations between microbes and diseases at large scale to prioritize candidate microbial biomarkers by integrating experimentally verified association data such as Disbiome.

Key Features:

• Multiplex Heterogeneous Network (MHEN): Constructs an MHEN by integrating Disbiome and other relevant biological databases to capture complex microbe–disease interactions.
• Graph Convolutional Neural Network (GCNN): Employs an end-to-end GCNN that integrates diverse prior biological knowledge to predict multiple association types in a single training session.
• Link Prediction Framework: Frames the task as link prediction on the MHEN to infer whether a microbe is reduced or elevated in association with a disease.
• Scalability and Efficiency: Uses experimentally verified association data instead of microbiome abundance profiles to enable large-scale prediction and prioritization of candidate microbes for experimental validation.
• Competitive Performance: Demonstrates competitive performance in large-scale cross-validation studies and case analyses compared to other approaches.

Scientific Applications:

• Biomarker discovery: Prioritizes potential microbial biomarkers associated with diseases for downstream experimental validation.
• Disease mechanism investigation: Supports investigation of the microbiome's role in disease mechanisms by predicting directional (reduced/elevated) microbe–disease associations.
• Experimental prioritization: Facilitates selection of candidate microbes for targeted experimental follow-up without generating new microbiome abundance data.

Methodology:

Constructs a multiplex heterogeneous network from Disbiome and other databases, applies an end-to-end graph convolutional neural network integrating prior biological knowledge, frames prediction as link prediction to infer reduced/elevated associations, and evaluates performance via large-scale cross-validation studies and case analyses.