Structure-Based Drug-Drug Interaction Detection via Expressive Graph Convolutional Networks and Deep Sets (Student Abstract)

Mengying Sun, Fei Wang, Olivier Elemento, Jiayu ZhouAAAI 2020 · DOI

Introduction

Adverse drug-drug interactions (DDIs) are a significant clinical concern, accounting for approximately 30% of adverse drug reactions. These interactions occur when the combined effect of two or more drugs is synergistic or antagonistic, sometimes leading to life-threatening outcomes. While deep learning models, particularly Graph Convolutional Networks (GCNs), have shown promise in predicting DDIs by analyzing molecular structures, many existing architectures lack the necessary expressive power to capture nuanced chemical features. Furthermore, a critical requirement for DDI prediction is permutation invariance: the predicted risk must remain identical regardless of the order in which the drug pair is processed.

Methodology

The proposed framework addresses these challenges by combining an expressive graph encoder with a symmetric decoder architecture.

Expressive Graph Convolutional Encoder (EPGCN)

Standard GCNs often rely on single-layer perceptrons (1LP) before neighborhood aggregation, which can limit their ability to distinguish between complex molecular neighborhoods. Drawing inspiration from Graph Isomorphism Networks (GIN), the authors implement an Expressive Graph Convolutional Encoder (EPGCN). This model utilizes Multi-Layer Perceptrons (MLPs) prior to aggregation and employs sum pooling instead of mean pooling. Sum pooling is theoretically more powerful as it preserves structural information, such as the total count of specific atoms, which mean pooling often obscures.

Deep Sets Decoder

To enforce permutation invariance, the framework utilizes a "Deep Sets" architecture for the interaction decoder. Two primary variants were explored:

  1. Invariant Decoder: This variant sums the latent representations of both drugs before applying non-linear transformations.
  2. Equivariant Decoder: This version uses neural network layers with specifically constrained weights to capture deeper interactions while maintaining mathematical symmetry.
The proposed framework architecture featuring a twin-branched GCN encoder and a Deep Sets interaction decoder.

The entire pipeline is structured as a "twin-branched" system, where each drug in a pair is processed by an identical GCN encoder before their features are merged in the decoder.

Experimental Setup

The model was evaluated using a real-world DDI dataset containing 1,317 unique adverse reactions and 63,473 unique drug pairs. For detailed analysis, six specific DDI types were selected (HL, DH, VD, BMT, TI, and AP), representing various frequencies within the dataset. Performance was measured using the Area Under the Curve (AUC) and accuracy across 5-fold cross-validation.

Results and Discussion

The "EPGCN-Deepset" models consistently outperformed baseline methods across all tested categories.

Predictive Performance

The EPGCN-Deepset-3 variant achieved notable AUC scores, reaching 0.926 for the HL category and 0.940 for DH. These results represent a significant improvement over standard GCN baselines, which typically achieved AUCs around 0.83.

Learning Efficiency and Expressivity

The inclusion of MLPs within the graph convolution layers allowed the model to map distinct chemical structures more effectively in the feature space. Training curves demonstrate that the expressive GCN (MLP-sum) converges to higher accuracy levels significantly faster than conventional 1LP-based models.

Training curves for the APH dataset comparing MLP-sum (expressive GCN) against 1LP-based baselines.

The use of sum pooling and MLP layers ensures that the network can distinguish between subtly different molecular environments, which is crucial for identifying the specific chemical triggers of adverse interactions.

Conclusion and Future Work

This research demonstrates that increasing the expressive power of graph encoders and enforcing symmetry through Deep Set decoders significantly enhances DDI prediction. Future work will focus on scaling the model to predict all 1,317 adverse reactions simultaneously in a multi-label setting. Additionally, improving model interpretability to identify specific functional groups responsible for interactions and analyzing the computational trade-offs of using deeper MLPs remain important areas for investigation.

Summary updated

BibTeX
@inproceedings{sun2020structure,
  author = {Mengying Sun and Fei Wang and Olivier Elemento and Jiayu Zhou},
  title = {{Structure-Based Drug-Drug Interaction Detection via Expressive Graph Convolutional Networks and Deep Sets (Student Abstract)}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2020},
  doi = {10.1609/aaai.v34i10.7236},
}
Figure