A multi-view feature fusion framework with interpretable graph convolution for predicting microbe-drug associations.

Area of Science:

• Computational biology
• Pharmacology
• Bioinformatics

Background:

• Predicting microbe-drug associations (MDA) is vital for drug development and precision medicine.
• Existing computational models often lack interpretability and struggle to identify key interaction mechanisms.
• Validating biological decision processes in current MDA models remains challenging.

Purpose of the Study:

• To propose IDEAL (Interpretability-Driven Evolvable Attentive Learning for Microbe-Drug Association), a novel multi-view framework for predicting MDA.
• To enhance the interpretability and biological validation of MDA predictions.
• To provide an evolvable and parameter-efficient approach for MDA modeling.

Main Methods:

• IDEAL integrates diverse data types: drug network topology, BERT-encoded drug semantics, drug fingerprints, microbe genome attributes, BERT-encoded microbe semantics, and metabolic pathways.
• GNNExplainer is employed within a graph convolutional network (GCN) pipeline to identify influential microbe-drug relations and refine graph structures.
• Elastic Weight Consolidation (EWC) is utilized for parameter-efficient adaptation and to prevent catastrophic forgetting as biological data evolve.

Main Results:

• The proposed IDEAL framework achieves strong predictive performance on three public datasets.
• IDEAL provides interpretable structural explanations for predicted microbe-drug associations.
• The method demonstrates a practical transfer-learning capability for evolving MDA benchmarks.

Conclusions:

• IDEAL offers a significant advancement in predicting microbe-drug associations by integrating multi-view data and enhancing interpretability.
• The framework provides biologically relevant insights into drug-microbe interactions.
• IDEAL presents a robust and adaptable solution for the growing field of precision medicine and drug discovery.