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

Yuanfei Dai, Chenhao Guo, Wenzhong Guo, Carsten EickhoffBriefings in Bioinformatics 2021 · DOI · OpenReview

TL;DR

This paper introduces a novel framework for predicting drug-drug interactions (DDIs) by using Wasserstein Adversarial Autoencoders (AAE) to generate more realistic "negative" drug interaction examples. This approach significantly improves the performance of knowledge graph embedding models for both link prediction and DDI classification tasks, outperforming existing methods.

What Problem Does This Paper Address?

Predicting drug-drug interactions (DDIs) is crucial for public health and patient safety, as unexpected interactions can lead to adverse drug reactions (ADRs). While knowledge graphs (KGs) are a powerful way to represent these interactions, training KG embedding models requires "negative" samples (interactions that don't occur). Traditional methods for generating these negative samples often use a uniform random approach, which produces simplistic examples that don't effectively train the model. This leads to models that are less robust and accurate in predicting real DDIs. The authors aim to address this by creating more sophisticated and informative negative samples.

Methods

The core of this work is a novel adversarial learning framework that leverages Wasserstein Adversarial Autoencoders (AAE). Here's a breakdown:

  • Adversarial Autoencoder for Negative Sampling: The autoencoder acts as a generator. It takes a positive DDI triplet (drug A, interaction, drug B) and, by removing one drug, tries to generate a plausible alternative drug that could have a similar interaction. This is done by learning a latent representation of drugs and interactions.
  • Gumbel-Softmax Relaxation: A key challenge in generating discrete data (like drug names) is the "non-differentiability" problem, which prevents gradients from flowing back to train the generator. The authors employ Gumbel-Softmax relaxation to create a differentiable approximation, allowing for stable training.
  • Wasserstein Distance: Instead of the standard Jensen-Shannon divergence, which can suffer from issues when distributions don't overlap, the authors use the Wasserstein distance. This metric is more robust for comparing distributions, especially when dealing with generated data that might not perfectly align with real data. It helps prevent model collapse and encourages more diverse generated samples.
  • Discriminator: A standard KG embedding discriminator is used, but it's trained on both the original positive triplets and the more realistic negative triplets generated by the AAE. This forces the discriminator to learn more nuanced representations of DDIs.

Datasets

The authors evaluated their framework on two widely used public datasets:

  • DeepDDI: Composed of 1,710 drugs and 86 interaction types from DrugBank, capturing 192,284 drug-drug pairs.
  • Decagon: Composed of 637 drugs and 200 interaction types from the TWOSIDES dataset, capturing 1,121,808 drug-drug pairs.

For both datasets, 80% of drug-drug pairs were used for training, 10% for validation, and 10% for testing.

Key Insights and Ideas

  • Plausible Negative Samples are Key: The most significant insight is that the quality of negative samples dramatically impacts KG embedding performance. Simply generating random negative interactions is insufficient; they need to be semantically similar enough to be "deceptive" to the model.
  • AAE as a Sophisticated Negative Sampler: Using an autoencoder to generate negative samples is a novel approach. It allows the model to learn drug similarities and interaction contexts, leading to more realistic and challenging negative examples.
  • Wasserstein Distance for Stability: The adoption of Wasserstein distance is crucial for stabilizing adversarial training, especially with discrete data, and preventing common issues like model collapse and lack of sample diversity.
  • Generalizability of the Framework: The proposed framework is designed to be plug-and-play, meaning it can be integrated with various existing KG embedding models (like ComplEx, SimplE, RotatE) to improve their performance without requiring fundamental changes to the base models.

Key Results

The proposed Wasserstein AAE-based framework consistently improved the performance of baseline KG embedding models on both datasets for both tasks:

  • Link Prediction: The framework achieved better performance across all metrics (MR, MRR, HITS@N%) compared to existing methods. For instance, the improved ComplEx model achieved the best results on both DeepDDI and Decagon datasets. On DeepDDI, the framework improved HITS@10% by an average of 3 points.
  • DDI Classification: Similar improvements were observed. The improved RotatE method yielded the best results on DeepDDI, while the improved SimplE method performed best on Decagon. On DeepDDI, the framework improved PR-AUC by an average of 2 points.
  • Qualitative Analysis: Visualizations of learned embeddings showed that drugs with similar effects or categories clustered together, supporting the model's ability to capture meaningful relationships. Table 4 also demonstrated that the generator samples more semantically relevant drugs as negative examples compared to random sampling.

Limitations and Future Work

  • Focus on Structure: The current model relies solely on the structure of the KG (triplets) and is agnostic to other drug properties or features. This limits its ability to leverage richer information.
  • Comparison with Clinical Trials: The paper focuses on improving KG embedding models and does not directly compare with clinical trials or traditional machine learning methods that might use more domain-specific features.
  • Future Work: The authors suggest extending this framework to other graph embedding tasks like recommendation, classification, and retrieval, especially for hierarchical data. They also note that exploring multi-class evaluation problems could lead to better end-to-end performance.

Summary updated

BibTeX
@article{dai2021drug,
  author = {Yuanfei Dai and Chenhao Guo and Wenzhong Guo and Carsten Eickhoff},
  title = {{Drug-drug interaction prediction with Wasserstein Adversarial Autoencoder-based knowledge graph embeddings}},
  journal = {Briefings in Bioinformatics},
  year = {2021},
  doi = {10.1093/bib/bbaa256},
}
Figure