Source-linked AI summary
MR-GNN: Multi-Resolution and Dual Graph Neural Network for Predicting Structured Entity Interactions
Nuo Xu, Pinghui Wang, Long Chen, Jing Tao, Junzhou Zhao
TL;DR
Predicting interactions between structured entities is important, but existing graph methods use fixed-size neighborhoods and encode each entity independently. MR-GNN combines multi-resolution features with dual graph-state LSTMs to summarize graphs and model pairwise interactions during feature extraction. The paper reports that it outperforms state-of-the-art methods on two benchmark datasets.
Problem
Existing graph methods use fixed-size receptive fields and extract features from each entity independently, despite interaction cues occurring across different-sized substructures.
Source: §1 · pp. 1–2Method
MR-GNN combines multi-resolution graph features with dual graph-state LSTMs that summarize each graph and extract pairwise interaction features during feature extraction.
Source: §1 · p. 2, Overview of our approach · p. 2Results
MR-GNN outperforms state-of-the-art methods on two benchmark datasets.
Source: Overview of our approach · p. 2Takeaways & Limitations
The authors report that MR-GNN can be extended to large graphs by assigning node weights to node groups based on node-degree distributions.
Source: §6 · p. 6
Abstract
from arXiv · showhide
Predicting interactions between structured entities lies at the core of numerous tasks such as drug regimen and new material design. In recent years, graph neural networks have become attractive. They represent structured entities as graphs and then extract features from each individual graph using graph convolution operations. However, these methods have some limitations: i) their networks only extract features from a fix-sized subgraph structure (i.e., a fix-sized receptive field) of each node, and ignore features in substructures of different sizes, and ii) features are extracted by considering each entity independently, which may not effectively reflect the interaction between two entities. To resolve these problems, we present MR-GNN, an end-to-end graph neural network with the following features: i) it uses a multi-resolution based architecture to extract node features from different neighborhoods of each node, and, ii) it uses dual graph-state long short-term memory networks (L-STMs) to summarize local features of each graph and extracts the interaction features between pairwise graphs. Experiments conducted on real-world datasets show that MR-GNN improves the prediction of state-of-the-art methods.
1 Introduction
Predicting interactions between structured entities matters in medicine and chemistry, but manual testing is costly and existing graph models can miss features at different scales or model interactions too late. MR-GNN combines multi-resolution graph features with dual graph-state LSTMs, and is reported to outperform state-of-the-art methods on two benchmark datasets.
1 Introduction
- Interaction prediction matters for drug regimens and material design, while testing every pair experimentally can bring potential patient harms and high time and monetary costs.The paper motivates prediction as a way to address the impracticality of examining all interactions through clinical or laboratory experiments.Source: §1 · p. 1
- A fixed receptive field can miss or overinclude substructures because relevant chemical groups may require different neighborhood sizes.The examples require three hops for Hydroquinone’s ArOH and two for Acetic acid’s COOH.Source: §1 · pp. 1–2
- Existing GCNs encode each entity independently and model interactions only at final prediction, although useful interaction features can occur at different convolution layers for different graphs.For the weak-acid example, the useful layer differs between Acetic acid and Hydroquinone.Source: §1 · p. 2
Overview of our approach.
- MR-GNN uses multiple receptive fields to extract local features and models pairwise interaction during feature extraction.Its Summary-LSTM aggregates multi-resolution features for each graph, while its Interaction-LSTM extracts features between graph pairs.Source: §1 · p. 2, Overview of our approach · p. 2
- MR-GNN outperforms state-of-the-art methods on two benchmark datasets.This is the reported result across the benchmarks.Source: Overview of our approach · p. 2
2 Problem Definition
The task is to predict an interaction label for an unseen pair of structured entities represented as graphs, using labeled graph pairs for supervised learning.
2 Problem Definition
- Each structured entity is represented as a graph G = (V, E), with nodes carrying c-dimensional feature vectors and neighbor sets defining node degrees.Node features may be latent embeddings or explicit attributes.Source: §2 · p. 2
- Entity interaction prediction is formulated as supervised learning from graph pairs and their corresponding labels to predict the label of an unseen pair.The label set contains k possible interactions.Source: §2 · p. 2
- The training data consists of q input pairs, each containing two entity graphs and a corresponding interaction label from the label set.Source: §2 · p. 2
3 Method
MR-GNN predicts structured-entity interactions by combining features from different receptive-field sizes with graph-state LSTMs that summarize individual graphs and model pairwise interactions. Its architecture uses weighted graph convolution layers, graph-gather layers, dual LSTMs, and fully connected prediction layers.
3.1 Overview
- MR-GNN combines weighted graph convolution layers, dual graph-state LSTMs, and fully connected layers to predict entity-interaction labels.The convolution layers extract features at different receptive-field sizes; the LSTMs summarize graph features and extract interaction features.Source: §3.1 · p. 2, Figure 3 · p. 3
3.2 Weighted graph convolution layers
- Degree-based weights distinguish center-node and neighbor transforms, repeated convolutions reach larger neighborhoods, and graph pooling takes elementwise neighborhood maxima.The third convolution iteration represents a three-hop neighborhood centered on the node.Source: §3.2 · p. 3
3.3 Graph-gather layers
- Graph-gather layers compute a weighted sum of node vectors at each convolution layer to produce a graph-state for that resolution.The first graph-state includes information from individual nodes.Source: §3.3 · p. 3
3.4 Dual graph-state lstms
- S-LSTM sequentially aggregates graph-states from smaller to larger receptive fields, while I-LSTM consumes paired graph-states to extract interaction features.The final S-LSTM hidden output is concatenated with global graph pooling to form each graph’s final state.Source: §3.4 · p. 4
3.5 Fully connected layers
- Fully connected layers predict interaction labels from the concatenated final graph representations and interaction features.The classifier applies linear layers with ReLU and softmax activations.Source: §3.5 · p. 4
4 Experiment
The experiments evaluate chemical interaction prediction as binary classification and drug interaction prediction as multi-class classification. MR-GNN outperforms comparison methods across the reported benchmarks, with additional tests of training-set size, efficiency, and parameter sensitivity.
4.1 Dataset
- The CCI task labels chemical pairs using score thresholds 900, 800, or 700 for positive samples and score 0 for negative samples; DDI has 86 interaction labels.CCI compounds are converted to graphs with 75-dimensional node features; DDI drugs are represented by SMILES strings and converted to graphs when possible.Source: §4.1 · p. 4, §4 · p. 4, Table 1 · p. 4
4.2 Baselines
- The comparisons include DeepCCI, DeepDDI, PIP, DGCNN, and SNR, spanning sequence-based, engineered-feature, and graph-based approaches.DeepCCI uses SMILES character sequences, DeepDDI uses structural similarity profiles, PIP uses fixed three-hop neighborhoods, DGCNN uses graph convolution, and SNR sums node features.Source: §4.2 · pp. 4–5
4.3 Binary classification
- 0.6%-2.5% accuracy, 0.6%-2.2% F1, 1.0%-2.3% recall, and 0.05%-0.32% AUC gains over DeepCCI; MR-GNN leads across the binary-classification metrics.The authors attribute the improvement to graph-based feature extraction and to multi-resolution receptive fields and graph-state LSTMs.Source: Table 2 · p. 5, §4.3 · p. 5
4.4 Multi-class classification
- MR-GNN leads on all five DDI metrics, improving AUPRC by 1.58%, micro average by 5.23%, macro recall by 5.46%, macro precision by 5.60%, and macro F1 by 1.58%.The reported ablations find that weighted GCLs contribute most, followed by S-LSTM and I-LSTM.Source: Table 3 · p. 5, §4.4 · p. 5
4.5 Efficiency and robustness
- Across CCI900 training-set proportions from 30% to 70%, MR-GNN always performs best; its training time is moderate among the compared methods.The authors describe its robustness as improving with larger training proportions and report that parameter variations change performance by less than 2%.Source: §4.5 · pp. 5–6, Figure 4 · p. 6
5 Related Work
The related work distinguishes node-level learning, single-graph prediction, and pairwise graph tasks. Prior pairwise graph methods largely target similarity or cross-graph links, while this paper studies universal graph interactions.
Node-level Applications
- Node-level applications include classification and link prediction, using skip-gram, autoencoder, and neighbor-aggregation embedding methods.Examples include DeepWalk and LINE [Perozzi et al., 2014], SDNE, and GCN and GraphSAGE [Defferrard et al., 2016; Thomas and Welling, 2017; Hamilton et al., 2017a].Source: §5 · p. 6
Single Graph Based Applications
- Single-graph applications commonly classify graphs or predict graph properties, producing one embedding per graph.Representations aggregate node embeddings by averages, sums, histograms, or CNNs using node ordering [Li et al., 2017; Duvenaud et al., 2015; Zhao et al., 2018; Hamilton et al., 2017b; Kearnes et al., 2016; Zhang et al., 2018a].Source: §5 · p. 6
Pairwise Graph Based Applications
- Pairwise graph neural-network work is described as limited, with prior studies focused mainly on graph similarity or links between nodes across graphs; this paper studies universal graph interactions.The cited examples are Bai et al. (2018) and Yanardag and Vishwanathan (2015) for similarity, and Alex et al. (2017) for cross-graph node links.Source: §5 · p. 6
6 Conclusion
The conclusion presents MR-GNN as a multi-resolution graph neural network for structured-entity interaction prediction. Experiments support its effectiveness, and the authors describe a degree-based route to extending it to large graphs.
Conclusion
- MR-GNN predicts structured-entity interactions by learning comprehensive features through a multi-resolution architecture.The authors report empirical analysis across different interaction-prediction tasks.Source: §6 · p. 6
- Experiments on different interaction-prediction tasks demonstrate the effectiveness of MR-GNN.Source: §6 · p. 6
- For large graphs, the authors propose assigning node weights to node groups based on the node-degree distribution; they plan to apply MR-GNN to other domains.Source: §6 · p. 6