DHENN: A Deeper Hybrid End-to-end Neural Network for Highly Accurate Drug-Drug Interaction Events Prediction

Di Wu, Wu Sun, Yi He, Zhong Chen, Xin Luo

Keywords: Drug-drug interactions, Knowledge graph, Graph neural networks, Deep neural networks

Abstract: Accurate prediction of drug-drug interactions (DDIs) is crucial for therapeutic safety yet poses a substantial challenge due to complex pharmacodynamics. Traditional DDI prediction methods often falter for three reasons. First, they simplify dependency structures among entities (e.g., drugs, targets, enzymes, and transporters) in bipartite networks, falling short in modeling their high-order interactions. Second, the over-smoothing effects constrain the depth of the adopted neural networks, thereby limiting their learning capacity. Third, they mainly decouple the stages of representation and prediction, leading to suboptimal solutions that overlook the potential for ground-truth DDIs to refine embedding generation. In response, this paper proposes Deeper Hybrid End-to-end Neural Network (DHENN), which integrates a Multimodal Knowledge Graph (MKG) with a Prediction-Enhanced Cascading Network (PECN) in an end-to-end learning manner. Specifically, MKG captures higher-order relationships across entities, offering a holistic view of DDIs. PECN mitigates over-smoothing by incorporating shallow embeddings into deeper layers, preserving node-level diversity. The endto- end learning manner guarantees that the representation learning and predictive modeling of MKG and PECN are formulated into a unified learning objective. Extensive experiments substantiate that DHENN outperforms thirteen competitors on two real-world DDI datasets.

Primary Area: learning on graphs and other geometries & topologies

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 4710