Protein--ligand interaction (PLI) prediction is critical in drug discovery,
aiding the identification and enhancement of molecules that effectively bind to
target proteins. Despite recent advances in deep learning-based PLI prediction,
developing a versatile model capable of accurate binding affinity scoring and
virtual screening in PLI prediction is an ongoing challenge. This is primarily
due to the lack of structure--affinity data, resulting in low model
generalization ability. We here propose a viable solution to this challenge by
introducing a novel data augmentation strategy along with a physics-informed
neural network. The resulting model exhibits significant improvement in both
scoring and screening capabilities. Its performance was compared to
task-specific deep learning-based PLI prediction models, confirming its
versatility. Notably, it even outperformed computationally expensive molecular
dynamics simulations as well as the other deep learning models in a derivative
benchmark while maintaining sufficiently high performance in virtual screening.
This underscores the potential of this approach in drug discovery,
demonstrating its applicability to both binding affinity scoring and virtual
screening.

通过引入新的数据增强策略和物理信息神经网络，我们提出了一种解决蛋白质 - 配体相互作用预测中结构 - 亲和力数据缺乏的问题的可行方法，并得到了显著的性能提升，适用于药物发现中的配体结合活性评分和虚拟筛选。