Authors
Guolongwei Dai, Tao Luo, Tao Lin, Lingjie Fan, Quan Wei, Fengyi Wang
Published in
iScience. Volume 29. Issue 11. Pages 118137. Nov 20, 2026. Epub Nov 06, 2026.
Abstract
Accurate prediction of drug-drug interactions (DDIs) is important for reducing risks associated with polypharmacy and adverse drug reactions. To overcome the limitations of experimental methods, we present AGSL-DDI, an end-to-end framework that learns interaction representations directly from SMILES sequences. Character-level embeddings are processed in parallel by a bidirectional long short-term memory network and a one-dimensional convolutional neural network to capture complementary global sequential dependencies and local chemical patterns. An adaptive gating mechanism dynamically integrates these features, while bidirectional multi-head cross-attention models fine-grained interactions between paired drugs. On the BioSNAP and DrugBank datasets, AGSL-DDI achieved AUROC values of 0.9852 and 0.9816 and AUPR values of 0.9837 and 0.9818, respectively. These results show that coordinated global-local representation learning and cross-drug interaction modeling can effectively characterize DDI patterns, providing a practical computational approach for medication risk screening and drug development.
PMID:
43034682
Bibliographic data and abstract were imported from PubMed on 15 Nov 2026.
Read full publication at:
Please sign in
to see all details.
Advertisement
Stats
- Recommendations n/a n/a positive of 0 vote(s)
- Views 6
- Comments 0