SimVec: predicting polypharmacy side effects for new drugs.

Area of Science:

• Pharmacology and Cheminformatics
• Computational Drug Discovery

Background:

• Polypharmacy, the use of multiple daily drugs, is effective for complex diseases but increases adverse drug reaction risks.
• Accurate prediction of polypharmacy side effects is vital for drug safety, particularly for novel therapeutics.
• Current knowledge graph (KG) methods struggle with new drugs due to limited KG connections.

Purpose of the Study:

• To address the limitations of existing KG-based approaches for polypharmacy side effect prediction in new drugs.
• To introduce SimVec, a novel method enhancing KG structure for improved prediction accuracy.
• To develop an effective learning process for drugs with sparse data.

Main Methods:

• SimVec enhances KG structure via structure-aware node initialization and weighted drug similarity edges.
• A novel 3-step iterative learning process updates node embeddings for side effects, similarity, and under-connected drugs.
• Evaluation of negative relation generation strategies, identifying cache-based approaches as optimal.

Main Results:

• The proposed SimVec method significantly outperforms existing KG-based models in polypharmacy side effect prediction.
• The enhanced KG structure and learning process effectively handle drugs with limited prior connections.
• Cache-based negative relation generation proved most effective for polypharmacy-related tasks.

Conclusions:

• SimVec offers a significant advancement in predicting polypharmacy side effects, especially for new drugs.
• The method's ability to leverage enhanced KG structures and iterative learning improves drug safety assessments.
• Optimized negative relation generation further refines the accuracy and applicability of polypharmacy prediction models.