DrugPred: an EdgeConv-GNN and Bio_ClinicalBERT based polypharmacy ADR prediction and specialist recommendation model

Area of Science:

• Pharmacogenomics and Computational Toxicology
• Artificial Intelligence in Healthcare

Background:

• Adverse drug reactions (ADRs) are a significant healthcare concern, particularly with polypharmacy.
• Existing methods often fail to predict side effects arising from drug-drug interactions (DDIs).
• There is a need for advanced computational approaches to identify combination-induced ADR risks.

Purpose of the Study:

• To propose a deep learning-based framework, DrugPred, for predicting ADR risks in multi-drug settings.
• To integrate individual drug effects, drug-drug interaction statistics, and biomedical text data.
• To enhance the accuracy and scalability of ADR prediction compared to traditional methods.

Main Methods:

• Phase 1: Multi-layer perceptron (MLP) to learn baseline drug-ADR association scores (OFFSIDES dataset).
• Phase 2: Graph Neural Network (GNN) with EdgeConv to model DDIs using interaction statistics (TwoSIDES dataset).
• Phase 3: Bio_ClinicalBERT for encoding drug pairs, combined with association and interaction scores via an attention mechanism for multi-label ADR prediction.

Main Results:

• The DrugPred framework achieved high predictive performance: 95.73% accuracy, 0.94 F1-score, 0.993 ROC-AUC, and 0.988 PR-AUC.
• The model demonstrated effectiveness in predicting ADR risks with high precision and recall.
• A retrieval-based guidance system was developed to map ADR risks to System Organ Classes (SOC) and medical specialists.

Conclusions:

• The DrugPred framework offers a scalable and effective approach for predicting ADRs in complex multi-drug scenarios.
• Integration of diverse data sources (drug effects, interactions, clinical text) significantly improves ADR prediction.
• The system provides valuable recommendations for clinical practice by linking predicted ADRs to relevant medical expertise.