FeatureDock: Protein-Ligand Docking Guided by Physicochemical Feature-Based Local Environment Learning using Transformer

Molecular docking, the task of predicting the binding structures between a protein and a small molecule ligand, plays a significant role in structural-based drug discovery. In recent years, numerous deep learning-based methods for molecular docking have emerged. State-of-the-art approaches such as DiffDock formulate the docking problem using diffusion generative models, exhibiting superior performance than traditional docking algorithms. However, despite the strong performance of these deep learning-based docking methods in predicting binding poses, they often lack a well-defined scoring function. This limitation poses challenges in effectively distinguishing between the strong and weak inhibitors during virtual screening. To address this limitation, we introduce FeatureDock, a transformer-based deep learning framework, which can accurately predict the protein-ligand binding poses as well as achieve a strong scoring power for virtual screening. FeatureDock extracts chemical features from local environments within protein structures and utilizes a Transformer encoder to predict probability density envelopes indicating where ligands are most likely to bind in the protein pocket. We also designed a scoring function, which encodes the predicted probability density envelope, to optimize and score the ligand poses. In addition, the attention mechanism in FeatureDock’s Transformer further enhances the model’s interpretability by providing the attention weights of each chemical feature from the protein structures in predicting the binding probabilities. When applied to virtual screening, we demonstrated that FeatureDock outperforms DiffDock, Smina and AutoDock Vina in distinguishing strong inhibitors from weak ones for both Cyclin-Dependent Kinase 2 (CDK2, an inactivated form) and Angiotensin-converting enzyme (ACE). The performance was assessed using Kullback–Leibler (KL) divergence and area under receiver operating characteristic (AUC) evaluation metrics. We also demonstrate that FeatureDock can accurately predict the binding poses, achieving an average RMSD of 2.4 Å when compared to CDK2-ligand co-crystal structures. We anticipate that our FeatureDock holds promise to be widely applied in virtual screening to assist in drug design. FeatureDock is available at https://github.com/xuhuihuang/featuredock.