Task-Specific Activity Cliff Prediction Method Based on Transfer Learning and a Hyper Connection Graph Model

Activity cliffs (ACs) are defined as significant changes in biological activity triggered by minor chemical structural modifications. Accurately predicting ACs is crucial for drug discovery and molecular optimization. Existing approaches often overlook the intricate structural relationships within compounds, limiting both predictive accuracy and generalizability. To address these challenges, we propose TS-AC, a novel task-specific activity cliff prediction framework that integrates transfer learning and a hyper connection graph architecture. Specifically, a large-scale drug-drug interaction (DDI) prediction task is constructed and pretrained to learn rich chemical knowledge, which is then transferred to activity cliff prediction to improve model generalization. In addition, a hyper connection graph module is designed to effectively represent and capture the interactions between the core and the substituent fragments in matched molecular pairs, enabling the model to better understand the impact of subtle structural modifications on biological activity. Comprehensive experiments on three data sets demonstrate that TS-AC consistently outperforms state-of-the-art methods. Visualization analyses further confirm the interpretability and rationality of the model's design.