[Paper Review] Drug-Drug Adverse Effect Prediction with Graph Co-Attention

Andreea Deac, Yu-Hsiang Huang+3 more|arXiv (Cornell University)|May 2, 2019

Computational Drug Discovery Methods43 references68 citations

TL;DR

The paper introduces MHCADDI, a graph neural network that jointly learns drug representations via intra-drug message passing and inter-drug co-attention to predict drug–drug adverse effects, achieving state-of-the-art AUROC on a large DDI dataset using only drug structures.

ABSTRACT

Complex or co-existing diseases are commonly treated using drug combinations, which can lead to higher risk of adverse side effects. The detection of polypharmacy side effects is usually done in Phase IV clinical trials, but there are still plenty which remain undiscovered when the drugs are put on the market. Such accidents have been affecting an increasing proportion of the population (15% in the US now) and it is thus of high interest to be able to predict the potential side effects as early as possible. Systematic combinatorial screening of possible drug-drug interactions (DDI) is challenging and expensive. However, the recent significant increases in data availability from pharmaceutical research and development efforts offer a novel paradigm for recovering relevant insights for DDI prediction. Accordingly, several recent approaches focus on curating massive DDI datasets (with millions of examples) and training machine learning models on them. Here we propose a neural network architecture able to set state-of-the-art results on this task---using the type of the side-effect and the molecular structure of the drugs alone---by leveraging a co-attentional mechanism. In particular, we show the importance of integrating joint information from the drug pairs early on when learning each drug's representation.

Motivation & Objective

• Motivate early prediction of polypharmacy side effects to reduce clinical and economic burden.
• Develop a neural architecture that learns joint drug representations from molecular graphs.
• Show that cross-drug co-attention improves prediction using only drug structures.
• Evaluate on a large DDI dataset and compare to existing baselines.
• Demonstrate ablations to highlight the importance of joint learning and attention mechanisms.

Proposed method

• Represent each drug as a molecular graph with atoms as nodes and bonds as edges.
• Apply T=3 message-passing blocks to each drug to learn inner drug representations.
• Incorporate co-attention across the two drugs to compute outer messages and form a joint representation.
• Use multi-head attention (K=8) to aggregate cross-drug information and update atom features.
• Readout by summing atom features to obtain drug-level vectors and score side-effect likelihoods.
• Train with margin-based ranking loss for binary classification or binary cross-entropy for multi-label classification.

Experimental results

Research questions

• RQ1 Can joint learning of drug representations via cross-drug co-attention improve DDI adverse effect prediction using only molecular structure?
• RQ2 How does integrating intra-drug message passing with inter-drug attention affect predictive performance across multiple side effects?
• RQ3 Do multi-head co-attention and early cross-modal information yield measurable gains over single-modality or late fusion baselines?
• RQ4 How does the proposed model compare to existing baselines on a large-scale DDI dataset in AUROC?

Key findings

• MHCADDI outperforms strong baselines including Decagon and Drug-Fingerprints on AUROC (0.882 vs. 0.872 for Decagon on the main binary setting).
• Ablation shows joint learning and early cross-drug information (co-attention) significantly improve performance over architectures that fuse drugs later or only use internal message passing.
• Multi-label variant (MHCADDI-ML) remains competitive, though slightly behind the single-label MHCADDI on AUROC.
• t-SNE analyses indicate discernible clustering of drug–drug embeddings per side effect, suggesting meaningful joint representations.