PaperlayerView exact revision on arXiv

Source-linked AI summary

Modeling polypharmacy side effects with graph convolutional networks

Marinka Zitnik, Monica Agrawal, Jure Leskovec

arXiv:1802.00543v2cs.LGq-bio.MNstat.ML

TL;DR

Identifying side effects caused specifically by drug combinations is difficult because interactions are complex and sparsely observed. Decagon uses a multimodal graph convolutional model for multirelational prediction, outperforming alternative approaches by up to 69% and achieving a 20% average gain.

  • Problem

    The paper addresses how to identify side effects specifically attributable to drug pairs rather than either drug alone.

  • Method

    Decagon constructs a multimodal graph of protein, drug-target, and drug-drug interactions and applies a graph convolutional model for multirelational edge prediction.

  • Results

    Decagon outperforms alternative approaches by up to 69% and achieves a 20% average gain in predictive performance.

  • Takeaways & Limitations

    The findings support jointly modeling side effects and incorporating drug-target and protein-interaction information in polypharmacy prediction.

  • Takeaways & Limitations

    When node features are unavailable, Decagon represents each node with a unique one-hot vector.

Abstract

from arXiv · show

The use of drug combinations, termed polypharmacy, is common to treat patients with complex diseases and co-existing conditions. However, a major consequence of polypharmacy is a much higher risk of adverse side effects for the patient. Polypharmacy side effects emerge because of drug-drug interactions, in which activity of one drug may change if taken with another drug. The knowledge of drug interactions is limited because these complex relationships are rare, and are usually not observed in relatively small clinical testing. Discovering polypharmacy side effects thus remains an important challenge with significant implications for patient mortality. Here, we present Decagon, an approach for modeling polypharmacy side effects. The approach constructs a multimodal graph of protein-protein interactions, drug-protein target interactions, and the polypharmacy side effects, which are represented as drug-drug interactions, where each side effect is an edge of a different type. Decagon is developed specifically to handle such multimodal graphs with a large number of edge types. Our approach develops a new graph convolutional neural network for multirelational link prediction in multimodal networks. Decagon predicts the exact side effect, if any, through which a given drug combination manifests clinically. Decagon accurately predicts polypharmacy side effects, outperforming baselines by up to 69%. We find that it automatically learns representations of side effects indicative of co-occurrence of polypharmacy in patients. Furthermore, Decagon models particularly well side effects with a strong molecular basis, while on predominantly non-molecular side effects, it achieves good performance because of effective sharing of model parameters across edge types. Decagon creates opportunities to use large pharmacogenomic and patient data to flag and prioritize side effects for follow-up analysis.

1 Introduction

Polypharmacy can improve treatment of complex diseases but substantially increases side-effect risk from difficult-to-identify drug–drug interactions. Decagon addresses the challenge by predicting precise side effects from a multimodal biological network and outperforms alternative approaches.

  • Motivation: Polypharmacy combines multiple medications to treat diseases resistant to single-drug activity, but increases side-effect risk through drug–drug interactions.These interactions are rare, difficult to test exhaustively, and often missed in small clinical studies.
  • Challenge: Precisely predicting which side effect a drug pair will cause is clinically important but had not been studied through predictive modeling.The task distinguishes side effects attributable specifically to the combination from effects caused by either drug alone.
  • Approach: Decagon constructs a two-layer multimodal graph of protein–protein, drug–protein, and side-effect-labeled drug–drug interactions for multirelational edge prediction.Each drug–drug edge type represents a distinct polypharmacy side effect.
  • Approach: Exploratory analysis finds that co-prescribed drugs share more target proteins than random pairs and that protein–protein interactions improve modeling of drug characteristics.These observations motivate integrating both drug-target and protein-interaction information.
  • Results: 69%: Decagon outperforms alternative approaches by up to 69% and achieves a 20% average gain in predictive performance.The comparison includes multirelational tensor factorization, graph representation learning, and adapted link-prediction methods.

2 Datasets

The study constructs a two-layer multimodal graph of drugs and proteins for multirelational polypharmacy side-effect prediction. It integrates human protein interactions with individual-drug, drug-combination, and drug-target data from multiple databases.

  • Network construction: The network contains drug and protein nodes, with 964 drug-drug edge types representing distinct side effects.The formulation treats polypharmacy side-effect identification as multirelational link prediction in a two-layer multimodal graph.
  • Protein-protein interactions: The human PPI network contains 19,085 proteins and 719,402 experimentally documented physical interactions.It is unweighted and undirected and integrates information from multiple published PPI resources.
  • Individual-drug side effects: SIDER contains 286,399 drug-side-effect associations across 1,556 drugs and 5,868 side effects.These associations were obtained by mining adverse events from drug-label text.
  • Individual-drug side effects: OFFSIDES adds 487,530 off-label associations spanning 1,332 drugs and 10,097 side effects.OFFSIDES was integrated with SIDER to describe side effects of individual drugs and drug combinations.
  • Polypharmacy side effects: TWOSIDES records 1,318 polypharmacy side-effect types across 63,473 drug combinations with effects greater than expected from either drug alone.The database was generated from adverse-event reporting systems.
  • Final integrated network: The linked final network has 645 drug and 19,085 protein nodes connected by 715,612 protein-protein, 4,651,131 drug-drug, and 18,596 drug-protein edges.These counts reflect the network after linking entity vocabularies across databases.

3 Data-driven motivation for Decagon approach

The graph’s side effects vary greatly in frequency, co-occur non-independently, and differ in their relationships with shared drug-target proteins. These observations motivate information sharing across side effects and integrating protein-protein interactions into Decagon.

  • 3 Data-driven motivation for Decagon approach: More than 53% of polypharmacy side effects occur in less than 3% of documented drug combinations, while frequent effects occur an order of magnitude more often.This frequency variation makes rare-side-effect prediction especially challenging and motivates sharing information across all side effects.
  • 3 Data-driven motivation for Decagon approach: Polypharmacy side effects co-occur non-independently in drug combinations, suggesting that joint modeling of multiple side effects can improve prediction.The paper quantifies co-occurrence by counting combinations and applying permutation testing against random co-occurrence.
  • 3 Data-driven motivation for Decagon approach: More than 68% of drug combinations share zero target proteins, highlighting the need for protein-protein interactions to connect proteins targeted by different drugs.Target-protein overlap is defined using Jaccard similarity between the target sets of the two drugs.
  • 3 Data-driven motivation for Decagon approach: Random drug pairs have smaller target-protein overlap than co-prescribed drugs, with p-value = 5e−120 under a 2-sample Kolmogorov-Smirnov (KS) test.The comparison is reported for random drug pairs versus co-prescribed drugs in Figure 2.
  • 3 Data-driven motivation for Decagon approach: The association between shared target proteins and side effects varies by side effect, with high blood pressure showing a stronger association than ribfracture.Figure 2 illustrates that this target-overlap trend is unequally observed across different side effects.

4 Graph convolutional Decagon approach

Decagon models polypharmacy side effects as multirelational link prediction on a multimodal graph of drugs, proteins, and side-effect relations. Its graph neural network encoder produces node embeddings, and a relation-aware decoder reconstructs labeled drug-pair edges.

  • Graph formulation: Decagon represents protein binding, drug–protein targeting, and particular polypharmacy side effects as labeled edges in a multimodal graph.The graph is defined as G = (V, R), with nodes such as proteins and drugs.
  • Prediction task: The prediction task identifies side-effect associations between drug pairs that cannot be attributed to either drug alone.For each drug pair, Decagon estimates whether an edge of side-effect type r belongs to the relation set R.
  • Relation-aware decoder: The decoder uses learned node embeddings and treats each edge label separately to score the likelihood of drug pairs interacting through each side-effect type.Decagon combines an encoder producing node embeddings with a tensor factorization decoder for modeling polypharmacy side effects.
  • Graph encoder: The encoder applies nonlinear, relation-specific graph convolutions that aggregate first-order neighbor features and can be stacked into multiple layers.If node features are unavailable, the first layer uses unique one-hot vectors for every node.

A C B

Decagon is an end-to-end graph convolutional model for multirelational link prediction that combines a relation-specific encoder with node-type-specific decoders. Its parameterization shares global drug-interaction structure across side effects while modeling relation-specific variation and polypharmacy interactions.

  • Encoder: The encoder aggregates neighboring hidden states separately for each relation, normalizes and sums them, then applies ReLU with shared parameters across graph nodes.This produces the next-layer hidden representation for each node in parallel.
  • Decoder: For each relation, the decoder combines pairs of node embeddings to score potential edges, with probabilities obtained by applying sigmoid to the score.Drug-pair relations use a dedicated decoder, while other node pairs use a bilinear form with relation-specific matrix Mr.
  • Decoder: Drug-pair decoding uses a global interaction matrix R modulated by side-effect-specific diagonal factors Dr, enabling shared structure across polypharmacy side effects.This design shares parameters across rare and frequent side effects and permits interactions between embedding dimensions.
  • Training: Decagon jointly trains relation-specific encoder weights Wr, relation-specific matrices Mr, global matrix R, and side-effect-specific matrices Dr in an end-to-end multimodal link-prediction model.Training uses cross-entropy with negative sampling and propagates gradients through both encoder and decoder.

5 Experimental setup

The experiments frame polypharmacy side-effect prediction as multirelational link prediction over drug pairs and evaluate Decagon against tensor, embedding, and feature-based baselines. Drug-pair associations are split by side-effect type, with model settings selected on validation data and performance measured using three ranking metrics.

  • Task formulation: Each drug pair may have zero, one, or multiple side-effect relation types, making prediction a multirelational link prediction task.The relation types correspond to all modeled polypharmacy side effects.
  • Data splitting: For each side-effect type, 80% of associated drug pairs train the model, while validation and test sets each contain 10%.The task predicts drug pairs associated with each side-effect type, with leakage controls described in the experimental setup.
  • Decagon model: Decagon assigns every drug pair–side-effect combination a probability and incorporates individual-drug side effects as node features x_i.The evaluation is designed to prevent circularity and information leakage when predicting true polypharmacy side effects.
  • Baselines: Decagon is compared with RESCAL and DEDICOM tensor decompositions, DeepWalk embeddings, and concatenated drug features.The baselines represent multirelational factorization, network embeddings, and drug-feature classifiers.
  • Model selection: Parameter settings are selected by validation-set grid search, while Decagon uses a 2-layer neural architecture with d^(1) = 64 and d^(2) = 32 hidden units.For non-multirelational methods, parameters are selected separately for each side-effect type; gradient boosting trees vary from 10 to 100 trees.
  • Evaluation: Performance is reported separately for each side-effect type using AUROC, AUPRC, and AP@50, with higher values indicating better performance.These metrics capture receiver-operating-characteristic area, precision-recall area, and average precision at 50.

6 Results

Decagon outperformed alternative approaches across 964 polypharmacy side-effect types and modeled molecularly grounded effects particularly well. Its highest-ranked novel predictions received literature support, while learned side-effect representations clustered according to co-occurrence patterns.

  • Comparative performance: 19.7% (AUROC), 22.0% (AUPRC), and 36.3% (AP@50) were Decagon’s improvements over alternative approaches across 964 side effect types.The evaluation used average performance values for 964 side effect types.
  • Performance by side-effect type: Decagon modeled side effects with strong apparent molecular underpinnings particularly well.This pattern was consistent with the predominantly pharmacogenomic information in Decagon’s multimodal graph.
  • Literature-based evaluation: Five out of ten highest-ranked novel predictions had literature evidence supporting both the drug pair and the predicted side effect.The predictions were specific and the supporting evidence was unlikely to arise from random selection of drug-pair and side-effect associations.
  • Side-effect representations: Side effects embedded close together in 2D space tended to co-occur in drug combinations.The representations were derived from the side-effect matrices Dr and visualized with t-SNE.
  • Side-effect representations: 1e−34 (p-value) showed that co-occurring or related side effects had more similar representations than expected by chance in a 2-sample KS test.Similarity was assessed using average Euclidean distances to the three most frequently co-occurring side effects.

7 Related work

Related work spans computational drug-combination prediction and graph neural networks. Decagon addresses gaps by predicting clinically relevant polypharmacy side effects and extending graph convolutions to multimodal graphs with many edge types.

  • Drug combination modeling: Prior computational pharmacology mainly modeled individual drugs or monotherapy, whereas this work considers drug combinations, or polypharmacy.These methods target drug–molecular-target associations, adverse drug reactions, and drug repurposing.
  • Drug combination modeling: n(n −1)/2 pairwise combinations arise from n drugs, making exhaustive experimental screening costly and time-consuming.Higher-order combinations create an even larger search space.
  • Drug combination modeling: For the first time, the approach predicts which, if any, polypharmacy side effects occur when multiple drugs are taken together by a patient.This provides a more direct path toward clinical translation than pointwise experimental interaction estimates.
  • Neural networks on graphs: The model extends graph convolutional networks with multiple edge types representing different side effects and efficient weight sharing for multimodal graphs.Graph convolutional networks had achieved state-of-the-art performance in social networks and knowledge graphs but had not yet been used for computational biology.

8 Conclusion

Decagon predicts side effects of drug pairs using a general graph convolutional neural network for large multimodal graphs with many relation types. The approach integrates molecular interaction networks with population-level patient side-effect data and motivates adding drug concentration information in future work.

  • Conclusion: Decagon predicts side effects associated with co-prescribed drug pairs.The paper presents it as an approach for predicting side effects of drug pairs.
  • Conclusion: Decagon is a general graph convolutional neural network for multimodal graphs with many relation types.Its design supports graphs in which nodes can be connected through a large number of different relation types.
  • Future work: The approach integrates protein-protein and drug-target networks with population-level patient side-effect data.Future work considers whether dosed drug concentration levels and other biomedical information could be integrated into the model.
Loading 1802.00543v2…