AffiGrapher: Contrastive Heterogeneous Graph Learning with Aromatic Virtual Nodes for RNA-Small Molecule Binding Affinity Prediction.

RNA molecules exhibit diverse structures and functions, making them promising drug targets. However, predicting RNA-small molecule binding affinity remains challenging due to limited experimental data and the structural variability introduced by multiple RNA conformations. To address these challenges, we propose AffiGrapher, a physics-driven graph neural network that integrates a physics-informed graph architecture with contrastive learning. Incorporating multiple RNA conformations allows the model to capture a wide range of structural information, enhancing prediction robustness. AffiGrapher achieves state-of-the-art performance in binding affinity prediction, as demonstrated in both 10-fold cross-validation with randomized splitting and 10-fold cross-validation with sequence homology splitting. In addition, we evaluate the model under the cold-start setting, where both RNAs and small molecules in the test set are unseen during training, and conduct a multiscenario evaluation under structural uncertainty. These experiments demonstrate the model's strong generalization ability even on predicted docking poses. Furthermore, the proposed method demonstrates exceptional potential in virtual screening tasks, highlighting its practical utility for RNA-targeted drug discovery. This study suggests that integrating physics-based architectures with contrastive learning may help address key challenges in RNA-small molecule affinity prediction.