Drug-drug interaction prediction with Wasserstein Adversarial Autoencoder-based knowledge graph embeddings.

Area of Science:

• Pharmacology
• Bioinformatics
• Artificial Intelligence

Background:

• Drug-drug interactions (DDIs) are critical for public health and drug development.
• Knowledge graph (KG) embedding methods are increasingly used for DDI prediction.
• Existing KG embedding methods suffer from simplistic negative sampling strategies.

Purpose of the Study:

• To propose a novel KG embedding framework for DDI tasks.
• To address limitations in negative sample generation for DDI prediction.
• To improve the accuracy of DDI link prediction and classification.

Main Methods:

• Developed a KG embedding framework utilizing adversarial autoencoders (AAEs).
• Incorporated Wasserstein distances and Gumbel-Softmax relaxation for stable training.
• Employed autoencoders for high-quality negative sample generation and plausible drug candidate identification.
• Used a discriminator to learn embeddings from positive and negative triplets.

Main Results:

• The proposed framework significantly improved performance on DDI link prediction and classification tasks.
• The method outperformed existing competitive baselines.
• AAEs generated higher quality negative samples, leading to more effective model training.

Conclusions:

• The AAE-based KG embedding framework offers a substantial advancement in DDI prediction.
• The approach effectively overcomes limitations of traditional negative sampling in DDI research.
• This work provides a more robust and accurate method for understanding and predicting drug-drug interactions.