Dockformer: A Transformer-Based Molecular Docking Paradigm for Large-Scale Virtual Screening

Molecular docking is a critical step in drug development, enabling the virtual screening of compound libraries to identify potential ligands for specific therapeutic target proteins. However, the computational cost of traditional docking methods scales proportionally with the size of the compound library, making high-throughput screening computationally prohibitive. Recently, deep learning (DL) has enabled the development of data-driven models to accelerate the docking process, yet few can simultaneously deliver high throughput and screening performance superior to that of traditional docking methods. To address this gap, a novel DL-based docking approach named Dockformer is introduced in this study. Dockformer integrates multimodal molecular representations, including 2-D graph topology and 3-D structural features, to encode both topological and geometric information, and predicts binding conformations along with confidence scores in an end-to-end framework. Experimental results demonstrate that Dockformer achieves docking success rates (root mean square deviation (RMSD) $\lt 2~\unicode{0x00C5}$ ) of 90.53% and 82.71% on the PDBbind core set and PoseBusters benchmarks, respectively. Compared to traditional docking methods such as AutoDock Vina and Glide, Dockformer achieves an inference speedup exceeding $100\times $ . Furthermore, it surpasses the best-performing DL competitor by an absolute improvement of 7.01% on the PoseBusters benchmark. In addition, the utility of Dockformer is validated in a practical virtual screening case study targeting SARS-CoV-2 main protease inhibitor discovery. Given its superior docking accuracy and computational efficiency, Dockformer represents a robust and generalizable framework for computer-aided drug design.