Comparative Learning Method for Drug Target Interaction Prediction Based on Heterogeneous Network

Abstract

Drug-Target Interaction (DTI) prediction serves as a pivotal task for accelerating drug repositioning and novel drug development, significantly reducing the time and costs of traditional experiments. Heterogeneous graph neural networks (hgnns) have become mainstream due to their capability to integrate multi-source biological data across "drug-target-disease-side effect" relationships. This paper examines the three-stage evolutionary trajectory of dti prediction methods using 2024's hncl-dti anchor papers as core references. It clarifies the characteristics of common datasets like hbn-a/b and evaluation metrics such as auc and aupr, while focusing on current bottlenecks including dataset scale bias, insufficient staticity, poor domain-specific model adaptation, and lack of interpretability. by comparing vertical domain solutions with general-domain approaches from the 2024-2025 ccf category a conference, the study evaluates the applicability of techniques like inductive learning and bias-corrected sampling. The core conclusion proposes integrating biomedical characteristics with general technologies to construct multi-source dynamic datasets and knowledge-enhanced hgnns. Future efforts should focus on knowledge-enhanced heterogeneous graph learning by deeply embedding biological knowledge (e.g., drug mechanisms and target annotations) into models, combining general-domain adaptive projection and attentional attribution techniques. This will establish a dual-dimensional evaluation system for "predictive performance and clinical applicability," ultimately achieving technical goals of "high precision, high interpretability, and high inductiveness" to provide more practical support for drug development.