همایش ملی بیوانفورماتیک ایران

صفحه اصلی / 4th international edition and 13th Iranian Conference on Bioinformatics

Enhancing Drug-Target Interaction Predictions through the Integration of Self-Organizing Maps and Graph-Based Representation Learning

نویسندگان :

Amir Mahdi Zhalefar¹ Hamid Khoeini² Zahra Narimani³

1- دانشگاه تحصیلات تکمیلی علوم پایه زنجان 2- دانشگاه تحصیلات تکمیلی علوم پایه زنجان 3- دانشگاه تحصیلات تکمیلی علوم پایه زنجان

کلمات کلیدی :

Drug-Target Interaction Prediction،Graph Representation Learning،Self-Organizing Map،Vector Quantization

چکیده :

Accurate prediction of drug-target interactions (DTIs) is crucial for accelerating drug discovery and therapeutic development. This task poses a significant challenge due to the complexity of drug-target relationships and the limitations of traditional experimental methods, which are often time-intensive and resource-demanding. GSRF-DTI (Zhu and Ning, 2024), a recently developed framework, integrates drug-target pair networks with graph representation learning to address these challenges. It constructs drug and target homogeneous networks and employs GraphSAGE for scalable node representation learning, followed by a random forest classifier for interaction prediction. GSRF-DTI has demonstrated robust performance across benchmark datasets, identifying both known and novel DTIs with high accuracy. This study enhances the GSRF-DTI framework by integrating an auxiliary extension that explores the incorporation of Self-Organizing Maps (SOM) (Pasa and Navarin, 2022) to improve predictive accuracy and biological interpretability. In this approach, features are derived from the drug-target pair network (DTPs-NET) using methods adapted from GSRF-DTI. GraphSAGE captures local neighborhood relationships within the network, while SOM provides an additional perspective by preserving global topological patterns and identifying biologically meaningful clusters. This integration of SOM resembles vector quantization techniques, encoding complex network patterns while maintaining computational efficiency. The combined feature representations are processed through dense layers, merged, and utilized by a Random Forest classifier, to make DTI predictions. Evaluations on benchmark datasets reveal that incorporating SOM yields an improvement in predictive performance compared to the baseline GSRF-DTI framework. Moreover, the use of SOM facilitates interpretability by mapping DTI data into biologically coherent clusters, potentially uncovering hidden relationships that may inform drug research. This combined approach underscores the value of integrating topological pattern recognition with graph-based learning for more accurate DTI predictions in computational drug discovery.