Fast polypharmacy side effect prediction using tensor factorization

Area of Science:

• Computational chemistry
• Pharmacology
• Bioinformatics

Background:

• Adverse drug reactions from combinations are a growing concern in medicine.
• Laboratory methods are insufficient for predicting these combinatorial effects.
• Computational approaches, including tensor factorization (TF), have shown potential but require optimization.

Purpose of the Study:

• To investigate the efficacy of optimized tensor factorization models for polypharmacy side effect prediction.
• To evaluate the performance and efficiency of TF models compared to existing methods.
• To determine the optimal incorporation of monopharmacy data within TF models.

Main Methods:

• Utilized tensor factorization (TF) models, specifically the SimplE model, for predicting polypharmacy side effects.
• Incorporated monopharmacy data as self-looping edges in a graph-based approach.
• Trained models using Python 3.8.12 with PyTorch 1.7.1 on NVIDIA GPUs.

Main Results:

• The SimplE TF model achieved state-of-the-art performance, with AUC ROC of 0.978, AUC PR of 0.971, and AP@50 of 1.000 across 963 side effects.
• The model reached 98.3% of its peak performance within two training epochs (approx. 4 minutes), demonstrating significant speed advantages.
• Integrating monopharmacy data as self-looping edges yielded slightly better results than using it for embedding initialization.

Conclusions:

• Optimized tensor factorization models, like SimplE, are highly effective for predicting polypharmacy side effects.
• These models offer a computationally efficient and accurate solution compared to existing methods.
• The study highlights the potential of TF for advancing drug safety and personalized medicine.