# Relation-aware graph structure embedding with co-contrastive learning for drug-drug interaction prediction

**Authors:** Mengying Jiang, Guizhong Liu, Biao Zhao, Yuanchao Su, Weiqiang Jin  
**arXiv:** [2307.01507](https://arxiv.org/abs/2307.01507)  
**Date:** 2023-07-04  
**Summary page:** https://www.summarizepaper.com/en/paper/2307.01507v2/

## Key points

- Introduces a novel method called RaGSECo for predicting multi-relational drug-drug interactions (DDIs)
- Existing methods struggle to effectively learn RaGSEs of new drugs, resulting in overfitting
- RaGSECo utilizes two heterogeneous drug graphs: multi-relational DDI graph and multi-attribute drug-drug similarity (DDS) graph
- Propagates RaGSEs of drugs from these graphs to ensure effective embeddings for all drugs
- Introduces co-contrastive learning module to enhance the learning process
- Learns representations of drug pairs (DPs) from interaction view and similarity view
- Collaborative supervision between views leads to more discriminative DP representations
- Evaluated through experiments on three different tasks using two real datasets
- Outperforms existing state-of-the-art prediction methods in terms of accuracy and performance
- Combines relation-aware graph structure embedding with co contrastive learning to improve prediction of multi-relational DDIs

## Summary

This paper introduces a novel method called RaGSECo for predicting multi-relational drug-drug interactions (DDIs). The existing methods in this field typically construct a multi-relational DDI graph and learn relation-aware graph structure embeddings (RaGSEs) of drugs from the graph. However, these approaches often struggle to effectively learn RaGSEs of new drugs, resulting in overfitting when testing DDIs involving such drugs. To address this limitation, the authors propose RaGSECo which utilizes two heterogeneous drug graphs: a multi-relational DDI graph and a multi-attribute drug-drug similarity (DDS) graph. These graphs are used to respectively learn and propagate the RaGSEs of drugs, ensuring that all drugs possess effective RaGSEs. Additionally, the authors introduce a co-contrastive learning module to enhance the learning process. This module learns representations of drug pairs (DPs) from two distinct views: interaction view and similarity view. By encouraging these views to collaboratively supervise each other, more discriminative DP representations are obtained. The effectiveness of RaGSECo is evaluated through experiments on three different tasks using two real datasets. The results demonstrate that RaGSECo outperforms existing state-of-the art prediction methods in terms of accuracy and performance. Overall, this paper presents an innovative approach that combines relation aware graph structure embedding with co contrastive learning to improve the prediction of multi relational DDIs. The proposed method addresses the limitations of existing approaches by enabling effective learning of RaGSEs for both known and new drugs.