A Review of Approaches for Predicting Drug–Drug Interactions Based on Machine Learning

Ke Han, Peigang Cao, Yu Wang, Fang Xie, Jiaqi Ma, Mengyao Yu, Jianchun Wang, Yaoqun Xu, Yu Zhang, Jie WanFrontiers in Pharmacology 2022 · DOI

TL;DR

This review provides a comprehensive roadmap of how machine learning is transforming the prediction of Drug-Drug Interactions (DDIs). By moving beyond slow, expensive laboratory experiments, these computational models leverage chemical similarity, biological networks, and natural language processing to identify potentially dangerous drug combinations before they reach the clinic.

What Problem Does This Paper Address?

As polypharmacy becomes the norm—with over 10% of the population taking five or more medications simultaneously—the risk of adverse drug reactions (ADRs) skyrockets. Drug-Drug Interactions (DDIs) occur when one drug alters the effect of another, potentially leading to decreased efficacy or life-threatening toxicity.

Traditionally, identifying these interactions relied on in vivo or in vitro testing, which is prohibitively expensive and cannot keep pace with the rapid development of new pharmaceuticals. The challenge lies in "predictive pharmacology": how can we use the massive amount of existing data on known drugs to accurately forecast how a brand-new compound will interact with the thousands of drugs already on the market?

Methods

The authors categorize the current machine learning landscape into several distinct "families" of approaches:

  • Traditional Similarity & Classification: These are the "baseline" methods. Similarity-based models operate on the "guilt by association" principle—if Drug A interacts with Drug X, then a drug chemically similar to A likely will too. Classification models treat DDI prediction as a standard binary task (interaction vs. no interaction).
  • Network Propagation: This treats the world of drugs as a massive social network.
  • Link Prediction: Uses graph topology to predict missing "edges" (interactions) between "nodes" (drugs).
  • Graph Embedding: Uses techniques like Graph Convolutional Networks (GCNs) to compress complex network structures into low-dimensional vectors that a model can easily process.
  • Matrix Factorization: This approach decomposes a large, sparse matrix of known interactions into smaller matrices representing latent features of drugs, then recombines them to fill in the blanks (similar to how Netflix predicts movie preferences).
  • Literature-Based Methods: A unique hybrid that uses Natural Language Processing (NLP) to extract interaction data from unstructured medical journals and then feeds that data into predictive models.
  • Deep Learning (DNNs): The paper highlights the rise of Deep Neural Networks, specifically multi-modal autoencoders that can integrate diverse data types (chemical structure, protein targets, and side effects) into a single predictive framework.

Datasets

The paper identifies the "Big Five" databases that fuel these models:

  1. DrugBank: The gold standard, containing detailed chemical, pharmacological, and pharmaceutical data.
  2. SIDER: Focuses on adverse drug reactions and side effects.
  3. TWOSIDES: Specifically designed for polypharmacy, documenting side effects seen only when drugs are used in combination.
  4. KEGG: Provides the biological context, mapping drugs to metabolic pathways and molecular interactions.
  5. PubChem: A massive repository of chemical structures and biological test results.

Key Insights and Ideas

  • The "Similarity" Foundation: Almost all successful models rely on some form of "similarity fusion." By combining chemical similarity (structure) with phenotypic similarity (side effects) and genomic similarity (shared targets), models become significantly more robust.
  • From Pairs to Networks: The most significant "aha moment" in the field is the shift from looking at two drugs in a vacuum to viewing them as part of a global biological network. Network-based methods capture "higher-order" relationships that simple chemical comparisons miss.
  • The Power of Latent Features: Matrix factorization and graph embeddings prove that the most important characteristics for predicting an interaction aren't always the ones we can see (like molecular weight), but rather "latent" mathematical features discovered by the algorithms themselves.

Key Results

While this is a review paper rather than a primary study, it highlights several high-performing models:

  • Decagon: A GCN-based model that can predict the specific type of side effect resulting from a DDI across a multi-modal graph.
  • DeepDDI: A gold-standard deep learning framework that uses structural similarity profiles to predict over 80 types of drug-drug and drug-food interactions with high accuracy.
  • Performance Shift: The authors note a clear trend where non-traditional methods (Graph Embeddings and Matrix Factorization) are consistently outperforming traditional similarity-based models in both precision and the ability to generalize to entirely new drugs.

Limitations and Future Work

The authors identify three critical frontiers for the next generation of DDI research:

  1. Beyond Binary Predictions: Most current models only predict if an interaction exists. Future work must focus on predicting the mechanism (e.g., "Drug A inhibits the metabolism of Drug B") and the severity.
  2. Multi-Drug Complexity: Most models look at drug pairs. In reality, elderly patients often take 10+ drugs. Predicting the "cocktail effect" of three or more drugs remains a massive computational challenge.
  3. Data Imbalance: There is a wealth of data on "positive" interactions (things that happen) but very little reported data on "negative" interactions (combinations that are proven safe). This imbalance can lead to models that are over-sensitive and prone to false positives.

Summary updated

Abstract

Drug-drug interactions play a vital role in drug research. However, they may also cause adverse reactions in patients, with serious consequences. Manual detection of drug-drug interactions is time-consuming and expensive, so it is urgent to use computer methods to solve the problem. There are two ways for computers to identify drug interactions: one is to identify known drug interactions, and the other is to predict unknown drug interactions. In this paper, we review the research progress of machine learning in predicting unknown drug interactions. Among these methods, the literature-based method is special because it combines the extraction method of DDI and the prediction method of DDI. We first introduce the common databases, then briefly describe each method, and summarize the advantages and disadvantages of some prediction models. Finally, we discuss the challenges and prospects of machine learning methods in predicting drug interactions. This review aims to provide useful guidance for interested researchers to further promote bioinformatics algorithms to predict DDI.
BibTeX
@inproceedings{han2022review,
  author = {Ke Han and Peigang Cao and Yu Wang and Fang Xie and Jiaqi Ma and Mengyao Yu and Jianchun Wang and Yaoqun Xu and Yu Zhang and Jie Wan},
  title = {{A Review of Approaches for Predicting Drug–Drug Interactions Based on Machine Learning}},
  booktitle = {Frontiers in Pharmacology},
  year = {2022},
  doi = {10.3389/fphar.2021.814858},
}
Figure