همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

Enhancing Drug Synergy Prediction through Relation and Structure Aware Attention Networks

نویسندگان :

Dorsa Abdi¹ Abdussalam Nik Kerdar² Parinaz Kanan³ Amirali Maroufi⁴ Maedeh Sadat Tahaei⁵

1- Iran University of Science and Technology Tehran, Iran 2- Iran University of Science and Technology Tehran, Iran 3- Iran University of Science and Technology Tehran, Iran 4- Iran University of Science and Technology Tehran, Iran 5- Iran University of Science and Technology Tehran, Iran

کلمات کلیدی :

drug synergy،attention mechanism،knowledge graph،heterogenous graph transformer،graph neural network

چکیده :

In the context of ever-more-complex biomedical data, drug discovery has had to evolve to be a data-driven science. A complex disease like cancer often cannot be treated successfully with a single agent due to the possibility of resistance or failure at the network level of interactions among agents, as well as the biological patient. Thus, it is a computationally difficult problem to predict drug synergy—the added therapeutic effect of a pair of drugs. Recently, KGANSynergy demonstrated the ability of knowledge graph attention networks to learn multi-relational biomedical dependencies. However, KGANSynergy is a general baseline model and is also limited in heterogeneous-oriented entity and relation encoding. In the present work, we build upon KGANSynergy in a comparative and heterogeneous-aware manner by implementing two alternative graph learning paradigms: (1) Structure-Aware Graph Transformer (SAGT) for type-subtyped message passing, and (2) Relation-Aware Graph Attention (RA-GAT) for fine-tuned relational modeling. Using publicly available benchmark datasets of drug-cell line interactions, we demonstrate how the use of heterogeneity-aware approaches improves predictive accuracy and interpretability to inform pre-clinical drug screening. Ultimately, our work shows how multi-relational learning is critical as we move closer to explainable and scalable drug combinations. In our study, we focused on the binary classification task of drug synergy prediction and achieved state-of-the-art performance, with an AUC of 0.8951 on the DrugCombDB dataset.