Capturing induced-fit effects: A geometry-aware and interpretable framework for robust drug-target affinity...

Accurate drug-target affinity (DTA) prediction is pivotal for virtual screening, yet practical reliability is often limited by the static treatment of molecular interactions and the "black-box" nature of deep learning models. In biological reality, binding involves dynamic structural adaptations known as induced-fit effects, which are frequently overlooked by conventional static representations. To address this, we propose a geometry-aware and interpretable framework (DCR-DTA) that explicitly models bidirectional induced-fit interactions, prioritizing stable structural anchors rather than computing raw 3D displacements. By integrating Dynamic Contextual Regularization, our model not only mitigates feature anisotropy in pretrained representations but also captures the intrinsic geometry of the interaction manifold. Extensive experiments on the Davis and KIBA benchmarks demonstrate that our method consistently outperforms state-of-the-art baselines, particularly in challenging cold-start scenarios. Notably, beyond achieving competitive Mean Squared Error (MSE), our model achieves superior external validation metrics, including a higher r_m² score (0.787 on KIBA) and a Concordance Index (CI) of 0.902. Crucially, visualization analyses reveal that the model learns discriminative interaction patterns that align with biological intuition, providing interpretable insights into the binding mechanism. These results suggest that explicitly modeling interaction dynamics and representation geometry is essential for robust and explainable DTA prediction. The source code and datasets are available at https://github.com/ycqmyxh/DCR-DTA.