Deep learning-based methods for drug target binding affinity (DTA) prediction are improving the efficien cy of drug screening, but some limitations persist in current methodologies. Notably, prevailing models predominantly rely on static structural data while neglecting the conformational dynamics of drug target complexes, which compromises their capacity to discern subtle conformation dependent affinity variations. To address this issue, we first constructed MD-PDBbind, an enhanced sampled molecular dynamics simulation (MD) dataset. Building upon this foundation, the MDDTA model incorporating the novel FAFormer architecture was proposed to achieve (3) equivariance and invariance, allowing the model to better learn the geo metric information of the drug target complexes. Further more, we formulated a dynamic-aware loss function to enhance the adaptability of model to diverse conformations. The MDDTA demonstrates excellent scoring and ranking performance on the CASF-2016 dataset, with a case study providing intuitive validation of the effectiveness of incor porating dynamic information. Lastly, a drug screening process was developed using the MDDTA to screen 70 SARS CoV-2 candidate compounds, five of which have been validated in the literature. These results highlight the potential of MDDTA for practical drug screening.
Zhao et al. (Thu,) studied this question.