Incorporating 3D Structural Information for Activity Cliff Analysis

Résumé fourni par la source

Artificial intelligence has become a crucial tool in drug discovery, excelling in tasks such as molecular property prediction. However, the activity cliff, where a minor structural modification to a molecule leads to a significant change in its biological activity, remains a challenge in predictive modeling. The activity cliff depends on the interaction between the target and the ligand, which is largely overlooked by previous ligand-centric studies. However, the limited availability of activity cliff data for target-ligand 3D complexes constrains the predictive power of modern deep learning models. To bridge this gap, we introduce DockedAC, a new dataset incorporating the protein target and 3D complex structure information to enable geometric reasoning in activity cliff analysis. By matching protein binding information and ligand bioactivity, we employ molecular docking to generate the complex structure for each activity value. The DockedAC dataset contains 82,836 activity data across 52 protein targets annotated with activity cliff information. This dataset represents a significant step toward large-scale activity cliff research using 3D complex structures. We conduct a comprehensive evaluation of structure-aware Graph Neural Networks against 2D baselines. Our experiments reveal that 3D structural information captures physical insights orthogonal to 2D descriptors. Notably, we demonstrate that integrating learned 3D geometric representations with topological features leads to significant improvements in predicting activity cliffs.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.