Source-linked AI summary
SumGNN: Multi-typed Drug Interaction Prediction via Efficient Knowledge Graph Summarization
Yue Yu, Kexin Huang, Chao Zhang, Lucas M. Glass, Jimeng Sun, Cao Xiao
TL;DR
DDI prediction must exploit large but noisy biomedical KGs while handling multi-typed interaction effects, a harder task than binary prediction. SumGNN extracts and summarizes relevant subgraphs, integrates KG and molecular information, and reports up to a 5.54% F1 improvement over the best baseline, particularly for low-resource relation types.
Problem
Large biomedical KGs are noisy and difficult to integrate with smaller curated datasets, while multi-typed DDI prediction remains harder and less developed than binary prediction.
Method
SumGNN extracts local drug-pair subgraphs, uses self-attention to prune reasoning pathways, and integrates summarized KG information with chemical fingerprints through multiple channels.
Results
Up to 5.54% increase over the best baseline in F1 is reported, with stronger results in low-resource relations, reduced inference time, and mechanism clues from reasoning paths.
Takeaways & Limitations
SumGNN supports multi-typed DDI prediction while using external biomedical knowledge and producing reasoning paths that provide clues about interaction mechanisms.
Abstract
from arXiv · showhide
Thanks to the increasing availability of drug-drug interactions (DDI) datasets and large biomedical knowledge graphs (KGs), accurate detection of adverse DDI using machine learning models becomes possible. However, it remains largely an open problem how to effectively utilize large and noisy biomedical KG for DDI detection. Due to its sheer size and amount of noise in KGs, it is often less beneficial to directly integrate KGs with other smaller but higher quality data (e.g., experimental data). Most of the existing approaches ignore KGs altogether. Some try to directly integrate KGs with other data via graph neural networks with limited success. Furthermore, most previous works focus on binary DDI prediction whereas the multi-typed DDI pharmacological effect prediction is a more meaningful but harder task. To fill the gaps, we propose a new method SumGNN: knowledge summarization graph neural network, which is enabled by a subgraph extraction module that can efficiently anchor on relevant subgraphs from a KG, a self-attention based subgraph summarization scheme to generate a reasoning path within the subgraph, and a multi-channel knowledge and data integration module that utilizes massive external biomedical knowledge for significantly improved multi-typed DDI predictions. SumGNN outperforms the best baseline by up to 5.54\%, and the performance gain is particularly significant in low data relation types. In addition, SumGNN provides interpretable prediction via the generated reasoning paths for each prediction.
1 Introduction
SumGNN addresses noisy, oversized biomedical KGs and the harder task of multi-typed DDI prediction by extracting relevant subgraphs, summarizing reasoning paths, and integrating multiple data channels. It improves prediction, especially for low-resource interaction types, while reducing inference time and offering mechanism clues.
- Curated DDI networks are high-quality but small and costly, whereas biomedical KGs are large and noisy, complicating effective integration for DDI prediction.
- SumGNN extracts local KG subgraphs around drug pairs to reduce noise and scale message passing through a smaller receptive field.
- A layer-independent self-attention scheme scores subgraph edges and prunes high-scoring pathways into sparse reasoning paths linked to biological processes.
- Multi-channel neural encoding combines summarized subgraph embeddings with chemical structures and other data sources for multi-typed DDI prediction.
- 5.54% increase over the best baseline in F1 is reported, alongside greatly reduced inference time and stronger performance in low-resource settings.
2 Related Works
Prior DDI approaches use KG embeddings or subgraph-based graph neural networks, but SumGNN adds knowledge summarization and multi-channel encoding for tractable, domain-specific pathway reasoning. The paper reports improvements over related approaches.
- Earlier DDI methods project entities and relations into dense KG embeddings, but do not directly use target-neighborhood information during inference.
- Graph neural networks and subgraph methods support relation modeling, graph learning, and scalability across tasks including link prediction and KG completion.
- SumGNN is presented as the first approach in this comparison to add graph summarization on subgraphs for tractable interaction pathways.
- SumGNN also introduces multi-channel neural encoding and reports improved predictive performance over the related works discussed.
3 Method
SumGNN predicts multi-relational drug interactions by anchoring on relevant KG subgraphs, summarizing them into sparse reasoning pathways, and integrating knowledge with molecular information. Its three modules reduce noise, support pathway-based interpretation, and produce drug-pair representations for classification.
- Local subgraph extraction: SumGNN extracts a local KG subgraph around each drug pair rather than processing the entire graph.This anchors computation on nearby biomedical entities, reducing noise and shrinking the message-passing receptive field.
- Multi-channel integration: The model initializes entity embeddings with TransE and aggregates representations across layers to generate a sufficient drug-pair representation.Layer aggregation concatenates node and subgraph embeddings from multiple GNN layers.
- Knowledge summarization: A layer-independent self-attention mechanism assigns edge signal intensities and prunes low-scoring edges to form a sparse interaction pathway.The attention scores use node representations and relation embeddings, then thresholding removes edges below γ.
- Knowledge summarization: Layer-independent pruning provides one unequivocal pathway for explanation and also reduces noise from potentially false-positive biomedical KG edges.Unlike layer-specific attention, the pruning decision depends only on the first-layer embedding.
- Multi-channel integration: SumGNN integrates summarized knowledge, subgraph features, and drug fingerprints through multi-channel neural encoding before decoding the interaction outcome.Morgan fingerprints provide molecular descriptors for drugs, while KG node types prevent using them as universal KG node features.
4 Experiments
Experiments evaluate multi-typed DDI prediction on DrugBank and TWOSIDES using curated train-development-test splits, repeated runs, and task-specific metrics. SumGNN is compared with molecular, network, GNN, and KG-integrated baselines.
- Datasets: DrugBank contains 1,709 drugs, 136,351 drug pairs, and 86 pharmacological relation types, while TWOSIDES contains 645 drugs, 46,221 pairs, and 200 side-effect labels.DrugBank is treated as multi-class after filtering multi-type edges; TWOSIDES remains multi-label with commonly occurring DDI types retained.
- Datasets: HetioNet supplies external biomedical knowledge from 29 public databases, with overlapping DDI edges removed to prevent information leakage.The resulting external KG contains 33,765 nodes across 11 entity or relation types, according to the dataset description.
- Baselines: SumGNN is compared with fingerprint MLPs, network embedding methods, DDI-network GNNs, and KG-based approaches including PRD, KG-DDI, GraIL, and KGNN.KGNN is identified as achieving state-of-the-art performance on binary DDI prediction, a task distinct from the multi-typed setting emphasized here.
- Metrics: DrugBank is evaluated as multi-class classification using accuracy, macro-averaged F1, and Cohen’s Kappa, while TWOSIDES uses averaged ROC-AUC, PR-AUC, and AP@50 across side-effect types.Macro-averaged class metrics give equal weight to classes, making them more sensitive to classes with fewer samples.
- Evaluation strategy: Both datasets use 7:1:2 train-development-test splits, with class coverage ensured for DrugBank and five independent runs conducted for every experiment.TWOSIDES additionally generates negative counterparts by sampling complements of positive examples.
4.2 SumGNN achieves superior predictive performance
SumGNN achieves the best reported multi-typed DDI prediction performance across two datasets, with especially strong gains for low-resource relations. Its subgraph-based summarization improves efficiency and produces sparse reasoning paths while parameter choices affect the balance between useful information and noise.
- Overall performance: SumGNN achieves the best DDI prediction performance on both DrugBank and TWOSIDES datasets.Table 1 reports averages and standard deviations over five runs, with higher metric values indicating better performance.
- Overall performance: 27.19%, 5.47%, and 4.65% absolute increases over the best baseline without KG are reported on three DrugBank metrics.On TWOSIDES, the corresponding increases are 2.84%, 2.45%, and 4.50%.
- Low-data relations: 57.14% F1 is achieved for relations with fewer than 10 training samples, while Decagon and KGNN make no correct predictions.The analysis groups relation types by training-set size and finds the largest advantage in extremely scarce-data settings.
- Overall performance: More than 4% average improvement on DrugBank and 2% on TWOSIDES over KGNN and KG-DDI shows that KG embeddings and neighborhood sampling alone are insufficient.The comparison supports the value of SumGNN’s subgraph-based knowledge use.
- Efficiency and interpretability: The summarization module prunes low-weight edges and retains sparse signal paths that provide clues about biological processes underlying drug interactions.A case study reports high attention weights for orthostatic hypotension and aplastic anaemia nodes.
- Parameter effects: Increasing the subgraph hop count can add useful information but may introduce noise, while overly large γ values can remove useful edges and reduce performance.For DrugBank, performance drops when k increases from 3 to 4; performance drops for γ > 0.4 on DrugBank and γ > 0.6 on TWOSIDES.
- Efficiency and interpretability: 80% of training time is saved through subgraph anchoring compared with propagating over the entire KG.The reduction follows from using local subgraphs rather than the full knowledge graph.
5 Conclusion
SumGNN combines local KG subgraphs, self-attention summarization, and multi-channel data integration for multi-typed DDI prediction. The framework performs well in low-resource settings and can extend to other interaction-prediction tasks.
- SumGNN uses local subgraphs, self-attention summarization, and multi-channel knowledge-data integration for multi-typed DDI prediction.
- SumGNN achieves good performance when training data for a relation type are scarce.The framework is described as adaptable to other interaction-prediction tasks.
- SumGNN can be adapted to predict interactions beyond drug-drug interactions, including drug-disease interactions.
A.1 SumGNN Parameter Setup
SumGNN uses fixed graph, feature, representation, and pruning settings selected through validation-set search, with key parameters examined experimentally.
- SumGNN uses 1,024-bit Morgan fingerprints for drug featurization and extracts 2-hop subgraphs.
- The subgraph summarization module uses d = 32 weight matrices and hidden representations.
- The relation matrix and input node representations are set to 32, while the edge-pruning threshold is γ = 0.
- The number of bases B in Eq. (3) is set to 8 because larger values can over-fit, while performance changes little from 4 to 16.
- The experiments study the effects of d, γ, and k.
A.2 Training Details
SumGNN and its baselines use shared training settings, including 50 epochs, batch size 256, ADAM optimization, regularization, and a two-layer GNN.
- Both SumGNN and the baselines are trained for 50 epochs with batch size 256.
- Training uses ADAM with learning rate 5 × 10^-3 and gradient clipping at 10 under the L2 norm.
- The model uses L2 weight decay of 1 × 10^-5, two GNN layers, and dropout rate 0.3 per layer.
- Implementations use PyTorch and Deep Graph Library on Ubuntu 18.04.3 with Python 3.6, PyTorch 1.2, and DGL 0.4.3.
A.3 The Range for Tuning Hyper-parameters
Key hyperparameters are selected with grid search, and their tuning ranges are summarized in Table 3, which marks default settings in bold.
- Grid search is used to determine hyperparameters.
- Table 3 reports the tuning ranges for key hyperparameters and identifies default settings with bold numbers.
A.4 Baseline Setup
The baseline setup uses specified implementations, architectures, dimensions, regularization, and training settings. Several models follow their authors’ released implementations or codebases, with KG embeddings trained via OpenKE for KG-DDI.
- MLP uses a two-layer Pytorch implementation with Morgan fingerprints, hidden dimension 100, and dropout 0.3.
- Node2vec follows the authors’ officially released implementation with embedding dimension 64.
- Decagon uses DGL with two GNN layers, hidden dimensions 64 and 32, dropout 0.1, and minibatch size 512.
- GAT uses DGL with hidden dimension 64, four attention heads, LeakyReLU activation, and α = 0.2; increasing heads hurts performance.
- SkipGNN, KG-DDI, GraIL, and KGNN use officially released or listed author codebases, while KG-DDI embeddings are trained with OpenKE.