Source-linked AI summary
GoGNN: Graph of Graphs Neural Network for Predicting Structured Entity Interactions
Hanchen Wang, Defu Lian, Ying Zhang, Lu Qin, Xuemin Lin
TL;DR
Structured-entity interaction prediction must account for both the structure of each entity and the interactions among entities, a gap in prior representations. GoGNN uses hierarchical graph neural networks with dual attention across local entity graphs and the global interaction graph, and outperforms state-of-the-art methods on chemical-chemical and drug-drug interaction tasks.
Problem
Structured-entity interaction prediction must capture both entity structure and interactions between entities, while prior works represent entities with insufficient information.
Source: §6 · p. 6Method
GoGNN hierarchically learns from local entity graphs and the global interaction graph, using dual attention to preserve the importance of substructures and interactions.
Source: §1 · p. 2Results
Takeaways & Limitations
The experiments support GoGNN's use for chemical-chemical and drug-drug interaction prediction, and the model can be naturally extended to other graph-of-graphs applications.
Source: §6 · p. 6
Abstract
from arXiv · showhide
Entity interaction prediction is essential in many important applications such as chemistry, biology, material science, and medical science. The problem becomes quite challenging when each entity is represented by a complex structure, namely structured entity, because two types of graphs are involved: local graphs for structured entities and a global graph to capture the interactions between structured entities. We observe that existing works on structured entity interaction prediction cannot properly exploit the unique graph of graphs model. In this paper, we propose a Graph of Graphs Neural Network, namely GoGNN, which extracts the features in both structured entity graphs and the entity interaction graph in a hierarchical way. We also propose the dual-attention mechanism that enables the model to preserve the neighbor importance in both levels of graphs. Extensive experiments on real-world datasets show that GoGNN outperforms the state-of-the-art methods on two representative structured entity interaction prediction tasks: chemical-chemical interaction prediction and drug-drug interaction prediction. Our code is available at Github.
1 Introduction
Structured-entity interactions matter across scientific and medical applications, but laboratory testing of all possible pairs is infeasible. GoGNN models entities as local graphs within a global interaction graph and applies dual attention across both levels; it outperforms state-of-the-art methods on chemical-chemical and drug-drug interaction prediction.
1 Introduction
- Interactions between structured entities support applications including toxicity prediction, material design, pollutant removal, drug discovery, and side-effect prediction.The paper notes that experimentally examining all possible interactions is infeasible in time and resources.Source: §1 · p. 1
- Existing approaches use only partial information: some omit entity graph structure, focus only on local structure or the interaction graph, or use no graph structure.MR-GNN [Xu et al., 2019] focuses on local structure and pairwise similarity, while Decagon [Zitnik et al., 2018] treats structured entities as simple nodes.Source: Figure 1 · p. 1
- The paper presents GoGNN as the first systematic application of graph neural networks on graph-of-graphs for structured entity interaction prediction.Source: §1 · p. 2
- GoGNN applies graph neural networks hierarchically to local entity graphs and the global interaction graph, with dual attention preserving substructure and interaction importance.The paper describes the two graph levels as synergistic.Source: §1 · p. 2
- GoGNN outperforms state-of-the-art methods on real-life chemical-chemical and drug-drug interaction prediction benchmarks.Source: §1 · p. 2
2 Related Work
Prior structured-entity interaction methods span conventional neural approaches and graph neural networks, while graph-level methods often omit interactions between graphs. GoGNN is situated as a graph-of-graphs approach for structured entity interaction prediction.
2 Related Work
- The paper introduces a section devoted to closely related works, establishing that related literature is discussed without specifying which research topics or methods the section covers.Source: §2 · p. 2
2.1 Structured Entities Interaction Prediction
- For chemical-chemical and drug-drug interaction prediction, prior methods include conventional CNN/PCA approaches and graph models such as Decagon, MR-GNN, and MLRDA.DeepCCI and DeepDDI use conventional convolutional neural networks and PCA; Decagon uses a drug-protein interaction graph, MR-GNN extracts local molecule features, and MLRDA uses a graph autoencoder.Source: §2.1 · p. 2
2.2 Graph Neural Networks
- Node-level GNNs model relations within a graph, while graph-level methods learn individual or pairwise graph representations without modeling interactions between graphs.Source: §2.2 · p. 2
2.3 Graph of Graphs
- Graph-of-graphs models treat individual networks as components of a larger multilevel network; the paper distinguishes GoGNN by targeting structured entity interaction prediction.The related-work passage describes GoGNN as the first GNN technique for this problem.Source: §2.3 · p. 2
3 Preliminaries
The paper formulates chemical-chemical and drug-drug interaction prediction as probability estimation over molecule graphs. It represents each molecule locally and their interactions globally, with edge attributes distinguishing drug side-effect types.
3.1 Problem Definition
- CCI prediction estimates the reaction probability for a chemical pair, whereas DDI prediction estimates the probability of a side-effect type for a drug pair.Source: §3.1 · p. 2
3.2 Input Graph of Graphs
- Each local molecule graph has atoms as nodes and bonds as edges, while the global interaction graph uses molecule graphs as nodes and interactions as edges.CCI has one interaction type; DDI has multiple side-effect types, represented by edge attribute vectors.Source: §3.2 · pp. 2–3
- Figure 2 illustrates the GoGNN framework.Source: Figure 2 · p. 3
4 Graph of Graphs Neural Network
GoGNN combines molecule-level graph processing with interaction-level neighbor aggregation to represent structured-entity interactions, using attention to select molecular substructures and weight interaction neighbors.
4.1 Framework of GoGNN
- Molecule-level features initialize interaction-level aggregation, whose feedback also helps the molecule network identify key substructures.The two GNN levels are described as synergistic.Source: §4.1 · p. 3
4.2 Molecule Graph Neural Network
- Stacked graph convolutions capture multi-hop molecular context, while attention pooling selects salient atoms and concatenated readouts retain multiresolution substructure information.The selected-atom embeddings are read out using mean and sum pooling.Source: §4.2 · pp. 3–4
4.3 Interaction Graph Neural Network
- Interaction-graph aggregation adds neighboring-molecule context, with attention weighting neighbors and edge-aware aggregation incorporating side-effect attributes for DDI.GoGNN shares edge-aggregation parameters across side-effect types.Source: §4.3 · p. 4
4.4 GoGNN Model Training
- For CCI, GoGNN treats prediction as link prediction, uses graph-representation dot products as link probabilities, and trains with randomly sampled non-edges.The negative sample pairs a positive-edge source molecule with a randomly chosen molecule.Source: §4.4 · p. 4
- For DDI, GoGNN models multirelational link prediction with cross-entropy training and negative samples formed by replacing the target drug according to a relation-specific distribution.The model uses side-effect-specific weights for the relation type.Source: §4.4 · p. 5
5 Experiment
Experiments evaluate GoGNN on chemical-chemical and drug-drug interaction prediction using real-world datasets, comparisons with established methods, ablations, and parameter sensitivity tests.
5.1 Dataset
- CCI900 and CCI950 contain 14,343 and 7,606 chemicals, while DDI and SE contain 1,704 and 645 drugs, respectively.The experiments use molecule graphs converted from SMILES strings, with 32-dimensional initial atom features and bond-type-weighted edges.Source: §5.1 · p. 5
5.2 Baselines
- The comparison includes CNN, MLP, GCN, multitask, and graph-neural-network baselines; the authors used public baseline code and reimplemented SEAL.The listed methods are DeepCCI, DeepDDI, Decagon, MR-GNN, MLRDA, and SEAL.Source: §5.2 · p. 5
5.3 CCI Prediction Results
- GoGNN outperforms all state-of-the-art baselines on CCI prediction, evaluated with AUC and average precision.The authors attribute the improvement to multiscale feature extraction and aggregation and dual attention over molecular substructures and chemical interactions.Source: §5.3 · p. 5, Table 1 · p. 6
5.4 DDI Prediction Results
- 1.18% and 1.19% gains on DDI AUC and AP, and 6.65% and 11.42% on SE AUC and AP; GoGNN also outperforms baselines on drug-interaction prediction.The reported gains are on the DDI and SE datasets, respectively.Source: Table 2 · p. 6, §5.4 · p. 6
5.5 Ablation Experiments
- Ablations identify GoGNN-M and GoGNN-I as the variants with the largest performance gaps, while graph-of-graphs, attention pooling, and attention-based and edge-filtered aggregation are effective.GoGNN-M omits interaction-graph convolution, GoGNN-I uses only the interaction graph, GoGNN-noPool replaces attention pooling, and GoGNN-noAttn replaces interaction-graph attention aggregation with a conventional GCN.Source: §5.2 · p. 5, §5.5 · p. 6
5.6 Parameter Sensitivity Analysis
- On CCI950, hyperparameter variation has insignificant overall impact; the reported best points are representation dimension 128, hidden feature size 384, learning rate 10 × 10−3, and pooling ratio 0.5.Source: §5.6 · p. 6, Figure 3 · p. 6
6 Conclusion
GoGNN addresses structured entity interaction prediction by using dual attention in a graph-of-graphs model. Experiments show improved performance on chemical-chemical and drug-drug interaction prediction; extension to other graph-of-graphs applications remains future work.
6 Conclusion
- GoGNN uses a graph-of-graphs view and dual attention to capture information from entity graphs and their interaction graph hierarchically.Source: §6 · p. 6
- Experiments on real-life datasets show improved performance for chemical-chemical and drug-drug interaction prediction.Source: §6 · p. 6
- Applying GoGNN to other graph-of-graphs settings, including financial and electrical networks, is left for future work.Source: §6 · p. 6