Drug-drug interaction prediction based on co-medication patterns and graph matching

Wen-Hao Chiang, Li Shen, Lang Li, Xia Ning

TL;DR

Researchers have developed a novel machine learning framework that predicts adverse drug reactions (ADRs) for drug combinations of any size, moving beyond simple pairs. By representing drug combinations as graphs and leveraging real-world "co-medication" patterns—how drugs are actually prescribed together—their model achieves a high predictive accuracy (AUC of 0.912) for identifying risks like myopathy.

What Problem Does This Paper Address?

Most research into Drug-Drug Interactions (DDIs) focuses on pairs—what happens when Drug A meets Drug B. However, the reality of modern medicine, especially for elderly patients, involves "polypharmacy," where individuals often take five or more medications simultaneously. These "high-order" interactions are significantly harder to predict because the number of possible combinations grows exponentially.

Existing methods often rely solely on the chemical structures of drugs. While useful, structure doesn't always tell the whole story of how a drug behaves in a complex biological system or why a physician might prescribe it alongside another. This paper seeks to bridge the gap by predicting ADRs for combinations of arbitrary orders using both structural data and clinical co-occurrence patterns.

Methods

The authors formulate the problem as a binary classification task using Support Vector Machines (SVMs). The core innovation lies in how they measure the similarity between two different drug combinations to feed the SVM.

  1. Graph Representation: Each drug combination is treated as a "complete graph" where nodes are drugs and edges represent the relationships between them.
  2. Graph Matching Kernels ($K_{gm}$): To compare two combinations (e.g., a 3-drug set vs. a 5-drug set), they use a graph matching algorithm that solves a Linear Sum Assignment Problem (LSAP). This finds the "best fit" between the drugs in one set and the drugs in another.
  3. Single Drug Similarity (SDS): To make the graph matching work, they need a way to say how similar two individual drugs are. They compared two approaches:
  4. $SDS_{2d}$: Traditional similarity based on 2D chemical structures (ECFP fingerprints).
  5. $SDS_{cm}$ (The Novel Approach): Similarity based on co-medication patterns. They look at how often drugs are prescribed together in patients who did and did not experience a specific ADR. The idea is that drugs used for similar therapeutic purposes or targeting similar pathways will show similar co-medication "signatures."

Datasets

The study utilized the FDA Adverse Event Reporting System (FAERS), a massive database of self-reported adverse events.

  • Target ADR: The researchers focused on myopathy (muscle disease) as their primary test case.
  • Data Volume: They extracted over 64,000 "case" events (myopathy reported) and nearly 1.5 million "control" events.
  • Benchmark: They also used the Side Effect Resource (SIDER) to validate which individual drugs are known to cause myopathy.
  • Pruned Dataset ($D^*$): A curated set of 1,210 drugs and several thousand combinations was used for the final model training and five-fold cross-validation.

Key Insights and Ideas

  • Clinical Wisdom as Data: The "aha moment" is using co-medication as a proxy for biological similarity. If two drugs are frequently used in similar clinical contexts, they likely share underlying mechanisms that chemical structure alone might not reveal.
  • Beyond Pairs: By using graph matching, the model isn't restricted to a fixed number of inputs. It can compare a "cocktail" of three drugs to one of six drugs seamlessly.
  • The "Statin" Signal: The model successfully identified that combinations involving statins (common cholesterol meds) were high-risk for myopathy, aligning with well-known medical literature, but it also found risks in combinations where no single drug was a known "offender."
  • Higher-Order Features Matter: The researchers found that "order-2" kernels (looking at pairs within the larger set) performed better than "order-1" kernels, suggesting that the interaction between drug pairs is a critical building block for understanding larger combinations.

Key Results

  • Top Performance: The graph matching kernel ($K_{gm}$) combined with co-medication similarity ($SDS_{cm}$) achieved an AUC of 0.912, outperforming traditional structural similarity and simpler "common drug" counting methods.
  • Accuracy: The best model reached an accuracy of 83.6%, a significant margin over baseline probabilistic kernels (76.5%).
  • Robustness: The supplementary materials show that the model remains consistent across different dataset sizes ($D^{2000}$ and $D^{4000}$), proving the approach isn't just a fluke of one specific data slice.

Limitations and Future Work

  • Single ADR Focus: While the results for myopathy are impressive, the paper only tests this one adverse reaction. Different ADRs might have different "signatures" that co-medication patterns may or may not capture as effectively.
  • Data Noise: FAERS data is self-reported and often incomplete or biased. The model's reliance on "frequency of use" could be skewed by how common a drug is in the general population rather than its actual risk profile.
  • Future Directions: The authors suggest moving toward Deep Learning to learn drug representations automatically from clinical data, potentially removing the need for pre-defined kernels and allowing the model to discover even more complex, non-linear drug relationships.

Summary updated

Abstract

Background: The problem of predicting whether a drug combination of arbitrary orders is likely to induce adverse drug reactions is considered in this manuscript. Methods: Novel kernels over drug combinations of arbitrary orders are developed within support vector machines for the prediction. Graph matching methods are used in the novel kernels to measure the similarities among drug combinations, in which drug co-medication patterns are leveraged to measure single drug similarities. Results: The experimental results on a real-world dataset demonstrated that the new kernels achieve an area under the curve (AUC) value 0.912 for the prediction problem. Conclusions: The new methods with drug co-medication based single drug similarities can accurately predict whether a drug combination is likely to induce adverse drug reactions of interest. Keywords: drug-drug interaction prediction; drug combination similarity; co-medication; graph matching
BibTeX
@article{chiang2020drug,
  author = {Wen-Hao Chiang and Li Shen and Lang Li and Xia Ning},
  title = {{Drug-drug interaction prediction based on co-medication patterns and graph matching}},
  journal = {Int. J. Comput. Biol. Drug Des.},
  year = {2020},
  url = {https://openreview.net/forum?id=1vIucAH5Ont},
}

Loading related content...

Figure