PaperlayerView exact revision on arXiv

Source-linked AI summary

Multi-view Graph Contrastive Representation Learning for Drug-Drug Interaction Prediction

Yingheng Wang, Yaosen Min, Xin Chen, Ji Wu

arXiv:2010.11711v3cs.LGcs.AI

TL;DR

DDI prediction is important because drug combinations can cause harmful effects, yet conventional detection is expensive and slow, and prior models incompletely combine molecular and interaction information. MIRACLE models DDI data as a multi-view graph, encodes molecular structure and DDI relationships with complementary graph networks, and uses contrastive learning to balance the views. Across multiple real datasets, it consistently outperforms state-of-the-art DDI prediction methods.

  • Problem

    DDI prediction is needed because drug combinations can alter effects, while wet-chemical experiments and clinical follow-up are costly and time-consuming and prior models may overlook important multi-view information.

  • Method

    MIRACLE treats each drug as a molecular graph within a DDI interaction graph, using bond-aware message passing, GCN integration, and contrastive learning to balance the two views.

  • Results

    MIRACLE consistently outperforms state-of-the-art DDI prediction models across multiple real-world datasets, including experiments with few labeled training instances.

  • Takeaways & Limitations

    Jointly modeling molecular structure and DDI-network relationships with balanced multi-view representations provides an effective and efficient DDI prediction approach.

  • Takeaways & Limitations

    Prior multi-view integration can blur molecular features through graph convolutions, while graph convolution may also cause over-smoothing and reduce structural information.

Abstract

from arXiv · show

Drug-drug interaction(DDI) prediction is an important task in the medical health machine learning community. This study presents a new method, multi-view graph contrastive representation learning for drug-drug interaction prediction, MIRACLE for brevity, to capture inter-view molecule structure and intra-view interactions between molecules simultaneously. MIRACLE treats a DDI network as a multi-view graph where each node in the interaction graph itself is a drug molecular graph instance. We use GCNs and bond-aware attentive message passing networks to encode DDI relationships and drug molecular graphs in the MIRACLE learning stage, respectively. Also, we propose a novel unsupervised contrastive learning component to balance and integrate the multi-view information. Comprehensive experiments on multiple real datasets show that MIRACLE outperforms the state-of-the-art DDI prediction models consistently.

1 INTRODUCTION

DDI prediction matters because drug combinations can alter effects and cause serious harm, while experimental and clinical detection is costly and slow. MIRACLE addresses limitations in prior approaches by jointly modeling molecular-graph and DDI-network views with contrastive learning, and experiments report consistent gains over state-of-the-art methods.

  • Motivation: Drug combinations can alter drug effects, increasing risks such as patient death or drug withdrawal.The passage reports a 20–40% prevalence in elderly patients and severe adverse drug reactions among hospitalized patients.
  • Motivation: Wet-chemical experiments and subsequent clinical testing are expensive, cumbersome, and time-consuming for DDI detection.These constraints motivate computationally assisted prediction methods.
  • Limitations of prior work: Prior similarity-based models require difficult feature engineering and may mispredict low-similarity interacting drug pairs.The passage also identifies missing features in large-scale datasets as a scalability and robustness problem.
  • MIRACLE: MIRACLE represents DDI data as a multi-view graph, treating each drug as a molecular graph and DDI relationships as an interaction graph.This formulation separates inter-view molecular structure from intra-view interaction relationships.
  • MIRACLE: MIRACLE encodes molecular graphs with bond-aware attentive message passing, integrates DDI relationships with a GCN, and balances views using contrastive learning.The learned drug embeddings are then used to construct interaction embeddings for prediction.
  • Results: MIRACLE evidently outperforms state-of-the-art methods across varied real-world datasets, including settings with few labeled training instances.The contribution summary reports consistent experimental superiority without specifying a numerical metric.

2 METHODOLOGY

MIRACLE is an end-to-end DDI prediction model with three sequential phases: molecular encoding, information integration, and interaction prediction. It uses bond-aware message passing for molecular graphs and a GCN to combine molecular embeddings with DDI-network connectivity.

  • Molecular encoding: MIRACLE first encodes drug molecular graphs into low-dimensional embeddings using bond-aware message passing with attentive pooling.The model uses atom and bond features together with molecular structure.
  • Information integration: The information integration module uses a GCN to combine inter-view drug embeddings with external DDI relationships.Drug instances are represented as nodes, embeddings as node features, and DDI connectivity as graph structure.
  • Interaction prediction: An interaction predictor uses the learned drug representations to predict missing interactions in the DDI network.The predictor is the final phase of the end-to-end pipeline.
  • Inputs and processing: The model accepts SMILES representations and a DDI-network connectivity matrix as inputs.SMILES strings are converted into molecular graphs before atom lists and multi-channel adjacency matrices are extracted.

2.1 Notations and Problem Formulation

The paper formulates DDI prediction as link prediction on a network whose nodes are drug molecular graphs. The goal is to predict missing interaction links.

  • A graph consists of vertices and edges, with atoms as vertices and chemical bonds as edges.
  • DDI prediction is defined as a link prediction problem on a graph.The network contains drug molecular graphs and interaction links.
  • Each drug is represented by a molecular graph, while the DDI network records interaction links between drugs.
  • The prediction task is to infer the existence of missing links in the DDI network.

2.2 Bond-aware Message Passing Networks with Attentive Pooling

MIRACLE encodes molecular graphs with bond-aware message passing and attentive pooling to produce low-dimensional drug embeddings. Its message updates distinguish chemical bond types and combine neighboring and previous node information.

  • Bond-aware message passing networks map molecular graphs to low-dimensional representations, followed by attentive pooling for graph-level embeddings.The molecular graph encoder processes atom and bond information before producing a drug representation.
  • Molecular graphs represent atoms as nodes and chemical bonds as edges, with separate adjacency channels for single, double, triple, and aromatic bonds.
  • The first message-passing phase shares affine-transformation parameters among neighboring nodes connected by the same chemical bond type.This design gives the message function a direct chemical interpretation.
  • The second update uses fuse, transform, and carry gates to combine candidate information with the previous hidden state.This accounts for both neighboring-node influence and the center node’s prior representation.
  • Attentive readout assigns atom-level importance scores and aggregates the weighted atom representations into a molecular-graph embedding.The resulting graph embedding has dimension d_g, and encoder parameters are shared across atoms.

2.3 GCN for Integrating Multi-view Network Information

MIRACLE integrates molecular and DDI-network information with a GCN encoder. The encoder takes the DDI adjacency and molecular-derived attributes as inputs and produces intra-view drug embeddings.

  • The GCN encoder integrates inter-view drug embeddings with external DDI relationships.
  • The DDI network is represented by an adjacency matrix A and an attribute matrix G for its drugs.
  • The adjacency matrix is normalized before graph convolution is applied.
  • The GCN produces an intra-view embedding matrix D for the drugs in the DDI network.

2.4 Contrastive Learning of Drug Representation

The contrastive component aligns two graph views by maximizing mutual information between inter-view molecular embeddings and intra-view network embeddings. It uses positive neighborhood pairs and negative samples across drugs to balance multi-view representations.

  • The DDI graph convolution can oversmooth node representations and reduce structural information.The molecular embeddings contain functional-group and local-connectivity information that graph convolutions may smooth.
  • MIRACLE maximizes agreement between inter-view and intra-view embeddings to learn drug representations.
  • For each anchor drug, positive samples include the drug itself and its k-order neighbors.
  • The method uses a mutual-information estimator implemented by a neural-network discriminator and optimized with a Jensen-Shannon-based lower bound.
  • Negative pairs are generated from all possible combinations of inter-view and intra-view embeddings across drugs in a batch.
  • Maximizing mutual information favors encoding information shared across different samples and views.

2.5 Drug-drug Interaction Prediction

MIRACLE combines supervised prediction with disagreement and contrastive losses to integrate molecular-graph and DDI-network views. Its graph pipeline encodes drug molecules and interaction links, with computational cost driven mainly by drug instances, chemical bonds, and DDI links.

  • Drug-pair embeddings are formed by element-wise multiplying the two drug vectors before a two-layer neural network predicts DDI occurrence.
  • An auxiliary predictor uses inter-view drug embeddings, allowing supervised information to flow directly into earlier molecular-graph layers.
  • The objective combines supervised, disagreement, and contrastive losses to align predictions and balance information across views.
  • The GNN computations require O(NC_u + Ld), where N is the number of drugs, C the chemical-bond count, L the DDI count, and u and d feature dimensions.
  • Full-batch gradient descent is used because the study’s DDI datasets contain far fewer than thousands of drugs and millions of interactions.

3 EXPERIMENTS

Experiments evaluate MIRACLE across small-, medium-, and large-scale DDI datasets under settings that vary feature availability and labeled-data scale. Across these evaluations, MIRACLE is reported to outperform baselines, including multi-view graph methods, while contrastive learning supports performance with fewer labeled interactions.

  • Datasets and settings: The evaluation uses ZhangDDI, ChCh-Miner, and DeepDDI as small-, medium-, and large-scale benchmarks with differing feature availability.ZhangDDI has 548 drugs, ChCh-Miner 1,514 drugs, and DeepDDI 192,284 pairwise DDIs.
  • Datasets and settings: The study removes DrugBank records containing improper SMILES strings that cannot be converted into molecular graphs.
  • Baselines and evaluation: MIRACLE is compared with similarity-based and graph-based baselines, using a shared architecture for graph methods that use different information views.
  • Comparative results: On ZhangDDI, combining similarity and topological features improves results for Ens, while single-feature similarity methods perform relatively poorly.
  • Comparative results: Multi-view graph baselines outperform single-view baselines, but MIRACLE achieves the best performance on datasets with few labeled DDI links.
  • Ablation and robustness: Contrastive learning improves drug representations and maintains high performance as the ZhangDDI training ratio decreases.
  • Comparative results: MIRACLE significantly outperforms other baselines on all three reported metrics in the evaluated comparison.The metrics are AUROC, AUPRC, and F1, reported as means and standard deviations over ten repetitions.
  • Convergence: The objective function decreases steadily after many iterations on the evaluated datasets.

4 RELATED WORK

DDI prediction has progressed from similarity-based features to graph representation learning and contrastive objectives, but existing approaches rely on restrictive similarity assumptions or limited graph views.

  • Overview: Existing DDI methods are commonly divided into similarity-based and graph-based approaches, with experimental assays remaining accurate but limited to small drug sets and environments.This motivates computational methods while framing the main methodological families reviewed in the paper.
  • Similarity-based methods: Similarity-based methods compare drugs using chemical, therapeutic, genomic, phenotypic, or other similarity features to generate or rank DDI candidates.These approaches commonly assume that drugs with similar chemical structures are prone to share similar interactions.
  • Graph-based methods: Graph-based methods learn higher-level representations from molecular, heterogeneous, or knowledge graphs for DDI prediction.Examples use graph convolutional networks, message passing, nearest neighbors, or knowledge-graph layers encoding local drug structures.
  • Graph representation learning: Traditional graph neural network research extends convolutional learning to non-Euclidean graph-structured data and supports tasks including link prediction and node classification.This broader GNN development provides the methodological foundation for graph-based DDI models.
  • Contrastive learning: Contrastive graph learning methods maximize agreement between representations from related graph views or distinguish original-graph nodes from corrupted-graph nodes.Prior work includes mutual-information maximization objectives for node- and graph-level representation learning.

5 CONCLUSION

MIRACLE integrates molecular-graph and DDI-network views through an end-to-end framework with contrastive learning and dual predictors. Experiments on real-life datasets demonstrate that it is effective and efficient.

  • 5 CONCLUSION: MIRACLE learns drug embeddings from multi-view DDI graphs using a bond-aware message passing network and a GCN encoder.The framework represents molecular structure and DDI relationships as complementary graph views.
  • 5 CONCLUSION: A contrastive learning strategy balances information from different views, while two predictors exploit information from both views.The design targets fuller integration of the available multi-view graph information.
  • 5 CONCLUSION: Extensive experiments on various real-life datasets demonstrate that MIRACLE is effective and efficient.The conclusion reports this result at the paper level without specifying a dataset or metric.
Loading 2010.11711v3…