ارزيابي روش‌هاي مبتني بر فرامسير براي پيش‌بيني ارتباطات دارو-هدف

eva‎luation Of Metapath Based Drug-Target Interaction Prediction Methods

رياضي كاربردي و علوم كامپيوتر

Drug repositioning, by reusing existing drugs for new therapeutic purposes, facilitates the faster discovery of therapeutic potentials. Recently, modeling biological data as heterogeneous graphs an‎d applying graph-based deep learning algorithms have attracted significant attention for drug–target interaction (DTI) prediction. However, the lack of precise distinction among biological entities an‎d relations has limited the accuracy an‎d generalization capability of existing models. To address these issues, metapaths have been utilized to enhance graph neural networks in capturing complex semantic an‎d structural information within heterogeneous graphs. Although metagraph-based approaches derived from metapath instances have demonstrated certain effectiveness, they still suffer from information leakage an‎d instability in prediction outcomes. Moreover, most biological databases lack explicit information on non-interacting pairs, which introduces ambiguity an‎d difficulty in model eva‎luation under the absence of negative samples. In this study, an analytical framework is proposed under conditions lacking negative samples, an‎d a hybrid strategy is designed to investigate challenges related to metagraph structures an‎d model generalization in discovering novel interactions. Two metapath-based an‎d metagraph-based models, MHGNN an‎d MHTAN-DTI, were examined an‎d eva‎luated. Experimental results revealed that strong performance in predicting known interactions does not necessarily indicate the ability to generalize toward new interaction discovery. It was also shown that incorporating metapaths containing direct drug–target edges may lead to information leakage. Additionally, constructing subgraphs from metagraphs based on interaction categories may cause fluctuations in prediction scores. Thus, the presented approach can serve as an effective tool for studying an‎d analyzing existing methods. Furthermore, to reduce the impact of metagraph structures on information leakage an‎d to address the challenge of lacking negative samples, suggestions for future research have been proposed that could pave the way toward developing more powerful, accurate, an‎d deployable models in the drug repositioning domain.

5